diff --git a/common/chat.cpp b/common/chat.cpp
index 50910242aedf..38a0139ce4f6 100644
--- a/common/chat.cpp
+++ b/common/chat.cpp
@@ -2123,6 +2123,189 @@ static common_chat_params common_chat_params_init_deepseek_v3_2(const common_cha
     return data;
 }
 
+// Kimi K3 - XTML-ish tagged format from the model's own template:
+//   open_tag(t, attrs) = <|open|>t k="v"...<|sep|>   close_tag(t) = <|close|>t<|sep|>
+//   assistant := [think] [response] [tools] close_tag(message) <|end_of_msg|>
+// Note the generation prompt already opens the think (or response) section, so
+// the section opener is optional here - same situation as Kimi K2 Thinking.
+static common_chat_params common_chat_params_init_kimi_k3(const common_chat_template &          tmpl,
+                                                          const autoparser::generation_params & inputs) {
+    common_chat_params data;
+
+    data.prompt            = common_chat_template_direct_apply_impl(tmpl, inputs);
+    data.generation_prompt = common_chat_template_generation_prompt_impl(tmpl, inputs);
+    data.format            = COMMON_CHAT_FORMAT_PEG_NATIVE;
+    data.supports_thinking = true;
+
+    const std::string SEP         = "<|sep|>";
+    const std::string MSG_START   = "<|open|>message role=\"assistant\"<|sep|>";
+    const std::string THINK_START = "<|open|>think<|sep|>";
+    const std::string THINK_END   = "<|close|>think<|sep|>";
+    const std::string RESP_START  = "<|open|>response<|sep|>";
+    const std::string RESP_END    = "<|close|>response<|sep|>";
+    const std::string TOOLS_START = "<|open|>tools<|sep|>";
+    const std::string TOOLS_END   = "<|close|>tools<|sep|>";
+    const std::string CALL_START  = "<|open|>call tool=\"";
+    const std::string CALL_END    = "<|close|>call<|sep|>";
+    const std::string ARG_START   = "<|open|>argument key=\"";
+    const std::string ARG_END     = "<|close|>argument<|sep|>";
+    const std::string MSG_END     = "<|close|>message<|sep|>";
+    const std::string EOM_TOKEN   = "<|end_of_msg|>";
+
+    // The four markers are the only special tokens; tag names ("think",
+    // "response", "message") are ordinary tokens and must NOT be preserved,
+    // or ordinary prose containing those words would be mangled.
+    data.preserved_tokens = {
+        "<|open|>",
+        "<|close|>",
+        "<|sep|>",
+        "<|end_of_msg|>",
+    };
+
+    data.thinking_start_tag = THINK_START;
+    data.thinking_end_tags  = { THINK_END };
+
+    // Per-role message-start delimiters. User/assistant messages carry only the
+    // role attribute, so their full opener (through <|sep|>) is used. System and
+    // tool messages continue with more attributes (type=/tool=/index=), so those
+    // delimiters stop after the role's closing quote - verified against the K3
+    // tokenizer that the quote is always its own token and never merges with the
+    // following attribute text, keeping the token-level prefix match exact.
+    data.message_delimiters = {
+        { COMMON_CHAT_ROLE_ASSISTANT, "<|open|>message role=\"assistant\"<|sep|>" },
+        { COMMON_CHAT_ROLE_USER,      "<|open|>message role=\"user\"<|sep|>"      },
+        { COMMON_CHAT_ROLE_TOOL,      "<|open|>message role=\"tool\""             },
+        { COMMON_CHAT_ROLE_SYSTEM,    "<|open|>message role=\"system\""           },
+    };
+
+    auto has_tools         = inputs.tools.is_array() && !inputs.tools.empty();
+    auto extract_reasoning = inputs.reasoning_format != COMMON_REASONING_FORMAT_NONE;
+    auto include_grammar   = has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE;
+
+    if (inputs.has_continuation()) {
+        const auto & msg = inputs.continue_msg;
+
+        data.generation_prompt = MSG_START + THINK_START + msg.reasoning_content;
+        if (inputs.continue_final_message == COMMON_CHAT_CONTINUATION_CONTENT) {
+            data.generation_prompt += THINK_END + RESP_START + msg.render_content();
+        }
+
+        data.prompt += data.generation_prompt;
+    }
+
+    auto parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
+        auto end = p.end();
+
+        auto start = p.optional(p.literal(MSG_START));
+
+        // The think section is ALWAYS consumed, even when reasoning extraction
+        // is off: K3's generation prompt ends with open_tag('think'), so the
+        // opener is present on every request and would otherwise leak into
+        // content. With extraction off the thoughts fall into content, matching
+        // how the other reasoning models behave.
+        // Reasoning stops at its own closer, or at the response opener if the
+        // model skips the closer entirely (seen on short answers).
+        auto think_body = extract_reasoning ? p.reasoning(p.until_one_of({ THINK_END, RESP_START })) :
+                                              p.content(p.until_one_of({ THINK_END, RESP_START }));
+
+        auto reasoning = p.optional(p.optional(p.literal(THINK_START)) + think_body +
+                                    p.optional(p.literal(THINK_END)));
+
+        // Content runs to the response closer, or to whatever comes next if a
+        // truncated generation never emits one.
+        auto response = p.optional(p.literal(RESP_START)) +
+                        p.content(p.until_one_of({ RESP_END, TOOLS_START, MSG_END })) +
+                        p.optional(p.literal(RESP_END));
+
+        // The message closer is followed by the EOG token, which reaches the
+        // parser as text and must be consumed or the parse is left incomplete.
+        auto trailer = p.optional(p.literal(MSG_END)) + p.optional(p.literal(EOM_TOKEN));
+
+        if (!has_tools || inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_NONE) {
+            return start + reasoning + response + trailer + end;
+        }
+
+        auto tool_choices = p.choice();
+        foreach_function(inputs.tools, [&](const json & tool) {
+            const auto & function = tool.at("function");
+            std::string  name     = function.at("name");
+            const json   schema   = function.contains("parameters") ? function.at("parameters") : json::object();
+
+            // Arguments arrive one tag per key, with the JSON type carried in a
+            // type="..." attribute. We take the type from the tool schema
+            // instead - it is authoritative, and it tells us whether the value
+            // should be parsed as JSON or kept as a literal string.
+            auto args = p.eps();
+            if (schema.contains("properties") && !schema.at("properties").empty()) {
+                auto arg_choices = p.choice();
+                for (const auto & prop : schema.at("properties").items()) {
+                    const std::string & key = prop.key();
+
+                    std::string type = "string";
+                    if (prop.value().is_object() && prop.value().contains("type") &&
+                        prop.value().at("type").is_string()) {
+                        type = prop.value().at("type").get<std::string>();
+                    }
+
+                    auto value = type == "string" ? p.tool_arg_string_value(p.until(ARG_END)) :
+                                                    p.tool_arg_value(p.until(ARG_END));
+
+                    // skip the trailing type="..." attribute: anything up to <|sep|>
+                    arg_choices |= p.rule("kimi-k3-arg-" + name + "-" + key,
+                                          p.tool_arg(p.tool_arg_open(p.literal(ARG_START)) +
+                                                     p.tool_arg_name(p.literal(key)) + p.literal("\"") +
+                                                     p.until(SEP) + p.literal(SEP) + value +
+                                                     p.tool_arg_close(p.literal(ARG_END))));
+                }
+                args = p.zero_or_more(arg_choices);
+            }
+
+            // skip the trailing index="N" attribute the same way
+            auto call = p.tool(p.tool_open(p.literal(CALL_START) + p.tool_name(p.literal(name)) + p.literal("\"") +
+                                           p.until(SEP) + p.literal(SEP)) +
+                               p.tool_args(args) + p.tool_close(p.literal(CALL_END)));
+
+            tool_choices |= p.rule("kimi-k3-tool-" + name, call);
+        });
+
+        // K3 emits every call inside one <|open|>tools<|sep|> section, then
+        // closes the message. The message closer is part of the trigger rule so
+        // that the lazy grammar still permits it once tool calls have started -
+        // otherwise constrained decoding rejects the model's own closing tag.
+        auto tools_section =
+            p.trigger_rule("kimi-k3-tool-call", p.literal(TOOLS_START) + p.one_or_more(tool_choices) +
+                                                    p.literal(TOOLS_END) + p.optional(p.literal(MSG_END)) +
+                                                    p.optional(p.literal(EOM_TOKEN)));
+
+        auto tools = inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED ? tools_section :
+                                                                              p.optional(tools_section);
+
+        return start + reasoning + response + tools + trailer + end;
+    });
+
+    data.parser = parser.save();
+
+    if (include_grammar) {
+        data.grammar_lazy = inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED;
+        data.grammar      = build_grammar([&](const common_grammar_builder & builder) {
+            foreach_function(inputs.tools, [&](const json & tool) {
+                const auto & function = tool.at("function");
+                if (function.contains("parameters")) {
+                    auto schema = function.at("parameters");
+                    builder.resolve_refs(schema);
+                }
+            });
+            parser.build_grammar(builder, data.grammar_lazy);
+        });
+
+        data.grammar_triggers = {
+            { COMMON_GRAMMAR_TRIGGER_TYPE_WORD, TOOLS_START },
+        };
+    }
+
+    return data;
+}
+
 // Cohere2 MoE (a.k.a. "North Code") parser.
 //
 // The assistant turn is fully marker-wrapped:
@@ -3057,6 +3240,14 @@ std::optional<common_chat_params> common_chat_try_specialized_template(
         return common_chat_params_init_kimi_k2(tmpl, params);
     }
 
+    // Kimi K3 - XTML-ish tagged format built from open_tag/close_tag macros.
+    // Detection: the <|open|>/<|close|>/<|sep|> marker trio is unique to K3.
+    if (src.find("<|open|>") != std::string::npos && src.find("<|close|>") != std::string::npos &&
+        src.find("<|end_of_msg|>") != std::string::npos) {
+        LOG_DBG("Using specialized template: Kimi K3\n");
+        return common_chat_params_init_kimi_k3(tmpl, params);
+    }
+
     // Cohere2 MoE / North Code - marker-wrapped format with <|START_TEXT|> content and
     // <|START_ACTION|> JSON tool calls. <|START_TEXT|> is unique to this template (the older
     // Command-R templates use <|START_RESPONSE|>).
diff --git a/conversion/__init__.py b/conversion/__init__.py
index abdf76cdba30..89f126f9227a 100644
--- a/conversion/__init__.py
+++ b/conversion/__init__.py
@@ -123,6 +123,7 @@
     "JinaEmbeddingsV5Model": "bert",
     "KORMoForCausalLM": "qwen",
     "KimiK25ForConditionalGeneration": "deepseek",
+    "KimiK3ForConditionalGeneration": "kimi_k3",
     "KimiLinearForCausalLM": "kimi_linear",
     "KimiLinearModel": "kimi_linear",
     "KimiVLForConditionalGeneration": "deepseek",
diff --git a/conversion/base.py b/conversion/base.py
index 720cfdca3a21..f938afa62e5f 100644
--- a/conversion/base.py
+++ b/conversion/base.py
@@ -72,6 +72,54 @@ class ModelType(IntEnum):
     MMPROJ = 2
 
 
+def repack_mxfp4_blocks(packed: Tensor, scale: Tensor) -> np.ndarray:
+    """
+    Repack 4-bit MX weights into ggml `block_mxfp4`. Lossless - this only moves
+    bits, it does not dequantize and requantize.
+
+    Source (compressed-tensors "mxfp4-pack-quantized", and DeepSeek-V4's
+    equivalent weight/scale pair):
+      packed  uint8 [rows, cols/2]  two 4-bit codes per byte, element 2i in the
+                                    low nibble and 2i+1 in the high nibble
+      scale   uint8 [rows, cols/32] one E8M0 biased exponent per 32-element group
+
+    Destination, per 32-element group: one scale byte then 16 code bytes, where
+    byte j holds element j in the low nibble and element j+16 in the high nibble
+    (see dequantize_row_mxfp4 in ggml-quants.c).
+
+    The 4-bit codes themselves need no remapping: both sides use sign in bit 3
+    and a magnitude index into (0, .5, 1, 1.5, 2, 3, 4, 6), which is exactly
+    ggml's kvalues_mxfp4 order. ggml's kvalues are doubled and its scale is
+    halved (GGML_E8M0_TO_FP32_HALF), so the represented value is unchanged.
+    """
+    p = packed.contiguous().view(torch.uint8)
+    s = scale.contiguous().view(torch.uint8)
+
+    rows, packed_cols = p.shape
+    cols = packed_cols * 2
+    if cols % 32 != 0:
+        raise ValueError(f"MXFP4 source row has {cols} values, expected a multiple of 32")
+
+    n_blocks = cols // 32
+    if tuple(s.shape) != (rows, n_blocks):
+        raise ValueError(f"MXFP4 scale shape {tuple(s.shape)} does not match {(rows, n_blocks)}")
+
+    # 0xff is the NaN E8M0 exponent; caught later without the tensor name
+    n_bad = int((s == 0xFF).sum())
+    if n_bad:
+        raise ValueError(f"invalid E8M0 scale byte 0xff in {n_bad} MXFP4 block(s)")
+
+    src = p.reshape(rows, n_blocks, 16)
+    lo = src & 0x0F           # elements 0, 2, 4, ...
+    hi = (src >> 4) & 0x0F    # elements 1, 3, 5, ...
+
+    vals = torch.stack((lo, hi), dim=-1).reshape(rows, n_blocks, 32)
+    qs = vals[:, :, :16] | (vals[:, :, 16:] << 4)
+
+    raw = torch.cat((s.unsqueeze(-1), qs.to(torch.uint8)), dim=-1)
+    return raw.reshape(rows, n_blocks * 17).cpu().numpy()
+
+
 class ModelBase:
     _model_classes: dict[ModelType, dict[str, type[ModelBase]]] = {
         ModelType.TEXT: {},
@@ -2664,7 +2712,10 @@ def get_model_architecture(hparams: dict[str, Any], model_type: ModelType) -> st
     # Step3-VL keeps text config under text_config but uses a custom top-level architecture.
     # For text conversion we route to a dedicated text-only class.
     # TODO: refactor this later to avoid adding exception here
-    if model_type == ModelType.TEXT and arch in ("StepVLForConditionalGeneration", "Sarashina2VisionForCausalLM", "Exaone4_5_ForConditionalGeneration", "Step3p7ForConditionalGeneration"):
+    # Kimi-K3's text_config reports "KimiLinearForCausalLM", which is the older
+    # Kimi-Linear-48B architecture and cannot load K3 (no attention residuals,
+    # latent MoE, situ, ...). Route on the top-level architecture instead.
+    if model_type == ModelType.TEXT and arch in ("StepVLForConditionalGeneration", "Sarashina2VisionForCausalLM", "Exaone4_5_ForConditionalGeneration", "Step3p7ForConditionalGeneration", "KimiK3ForConditionalGeneration"):
         return arch
 
     # if "architectures" is found in the sub-config, use that instead
diff --git a/conversion/deepseek.py b/conversion/deepseek.py
index ea6ae23d58e7..906aad67caff 100644
--- a/conversion/deepseek.py
+++ b/conversion/deepseek.py
@@ -12,7 +12,7 @@
 if TYPE_CHECKING:
     from torch import Tensor
 
-from .base import LazyTorchTensor, MmprojModel, ModelBase, TextModel, gguf, logger
+from .base import LazyTorchTensor, MmprojModel, ModelBase, TextModel, gguf, logger, repack_mxfp4_blocks
 
 from .qwen import QwenModel
 
@@ -596,30 +596,7 @@ def dequant_fp8_weight(weight: Tensor, scale: Tensor) -> Tensor:
         for name in tensors_to_remove:
             del self.model_tensors[name]
 
-    @staticmethod
-    def _pack_mxfp4_blocks(weight: Tensor, scale: Tensor) -> np.ndarray:
-        packed = weight.contiguous().view(torch.uint8)
-        scale_u8 = scale.contiguous().view(torch.uint8)
-
-        out_features, packed_cols = packed.shape
-        logical_cols = packed_cols * 2
-        if logical_cols % 32 != 0:
-            raise ValueError(f"MXFP4 source row has {logical_cols} values, expected a multiple of 32")
-
-        n_blocks = logical_cols // 32
-        if tuple(scale_u8.shape) != (out_features, n_blocks):
-            raise ValueError(f"MXFP4 scale shape {tuple(scale_u8.shape)} does not match {(out_features, n_blocks)}")
-
-        src = packed.reshape(out_features, n_blocks, 16)
-        low = src & 0x0F
-        high = (src >> 4) & 0x0F
-
-        # The safetensors bytes store adjacent values as low/high nibbles.
-        # ggml MXFP4 blocks store values 0..15 in low nibbles and 16..31 in high nibbles.
-        vals = torch.stack((low, high), dim=-1).reshape(out_features, n_blocks, 32)
-        qs = vals[:, :, :16] | (vals[:, :, 16:] << 4)
-        raw = torch.cat((scale_u8.unsqueeze(-1), qs.to(torch.uint8)), dim=-1)
-        return raw.reshape(out_features, n_blocks * 17).cpu().numpy()
+    _pack_mxfp4_blocks = staticmethod(repack_mxfp4_blocks)
 
     def _write_mxfp4_expert_tensor(self, bid: int, proj: str, tensor_key: gguf.MODEL_TENSOR) -> list[str]:
         n_experts = self.hparams["n_routed_experts"]
diff --git a/conversion/kimi_k3.py b/conversion/kimi_k3.py
new file mode 100644
index 000000000000..1064249e5205
--- /dev/null
+++ b/conversion/kimi_k3.py
@@ -0,0 +1,390 @@
+from __future__ import annotations
+
+import re
+from typing import Callable, Iterable, Iterator, TYPE_CHECKING
+
+import numpy as np
+import torch
+
+if TYPE_CHECKING:
+    from torch import Tensor
+
+from .base import LazyTorchTensor, ModelBase, TextModel, gguf, logger, repack_mxfp4_blocks
+
+from .kimi_linear import KimiLinearModel
+
+
+@ModelBase.register("KimiK3ForConditionalGeneration")
+class KimiK3Model(TextModel):
+    """
+    Kimi-K3 text model (KimiLinearForCausalLM under a `language_model.` prefix).
+
+    Shares the hybrid MLA + KDA skeleton with Kimi-Linear-48B but is not
+    loadable by that converter: K3 adds cross-layer attention residuals, a
+    latent MoE, the situ activation, an MLA output gate and a full-rank KDA
+    gate, none of which exist in the older architecture.
+
+    The vision tower and mm_projector are skipped - text only for now.
+    """
+
+    model_arch = gguf.MODEL_ARCH.KIMI_K3
+
+    _experts: list[dict[str, Tensor]] | None = None
+
+    # `<x>_res_norm.weight` and `<x>_res_proj.weight` are only ever used as the
+    # elementwise product norm.weight * proj.weight (see _apply_attn_res in
+    # modeling_kimi_linear.py), so they are fused into a single [n_embd] vector
+    # at conversion time. They arrive as separate tensors, so buffer whichever
+    # comes first, tagged with which one it is.
+    _res_parts: dict[str, tuple[str, Tensor]]
+
+    # HF suffix -> (gguf tensor, per-layer?)
+    _RES_FUSIONS = {
+        "self_attention_res": (gguf.MODEL_TENSOR.ATTN_RES_SCORE, True),
+        "mlp_res":            (gguf.MODEL_TENSOR.FFN_RES_SCORE,   True),
+        "output_attn_res":    (gguf.MODEL_TENSOR.OUTPUT_RES_SCORE, False),
+    }
+
+    # compressed-tensors MXFP4. The `language_model.` prefix is still present here:
+    # self.model_tensors is keyed by the raw checkpoint names, get_tensors() strips
+    # the prefix only on the way out.
+    _MXFP4_FORMAT = "mxfp4-pack-quantized"
+    _MXFP4_EXPERT_RE = re.compile(
+        r"^(?:language_model\.)?model\.layers\.(\d+)"
+        r"\.block_sparse_moe\.experts\.(\d+)\.(w[123])\.weight_packed$"
+    )
+    _MXFP4_PROJ = {
+        "w1": gguf.MODEL_TENSOR.FFN_GATE_EXP,
+        "w2": gguf.MODEL_TENSOR.FFN_DOWN_EXP,
+        "w3": gguf.MODEL_TENSOR.FFN_UP_EXP,
+    }
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self._res_parts = {}
+
+    def set_vocab(self):
+        # K3 ships the same TikToken vocab as K2: its pre-tokenizer hashes to
+        # 81212dc7... which base.py already maps to "kimi-k2", so no new
+        # pre-tokenizer registration is needed.
+        #
+        # Borrowed rather than inherited: K3 shares kimi-linear's vocab handling
+        # but none of its tensor layout. The method only touches TextModel
+        # members, so an unrelated TextModel is a valid receiver.
+        KimiLinearModel.set_vocab(self)  # ty: ignore[invalid-argument-type]
+
+        # ...but K2's converter ends by forcing eos to the tokenizer's own
+        # eos_id, which for K3 is 163585 = [EOS], the *document* terminator.
+        # K3's config and generation_config both say 163586 = <|end_of_msg|>,
+        # the chat turn terminator. Keeping [EOS] means generation never stops
+        # at the end of an assistant turn. Restore the configured value.
+        if (eos := self.hparams.get("eos_token_id")) is not None:
+            logger.info(f"restoring configured eos_token_id {eos} (kimi-linear forces the tokenizer's)")
+            self.gguf_writer.add_eos_token_id(eos)
+
+    #
+    # compressed-tensors MXFP4 -> ggml MXFP4
+    #
+    # The real checkpoint stores only the routed experts quantized (everything
+    # else is excluded by quantization_config["ignore"]), as a
+    # weight_packed/weight_scale pair per expert. Both sides are 4-bit E2M1 with
+    # a per-32 E8M0 scale, so this is a pure repack - see repack_mxfp4_blocks.
+    #
+    # Dequantizing instead would be catastrophic here: the routed experts are
+    # ~1.38 TB at 4 bits, so a bf16 round-trip would need ~5.5 TB of output.
+    #
+
+    def _is_mxfp4_packed(self) -> bool:
+        quant_config = self.hparams.get("quantization_config") or {}
+        return (quant_config.get("quant_method") == "compressed-tensors"
+                and quant_config.get("format") == self._MXFP4_FORMAT)
+
+    def dequant_model(self):
+        if not self._is_mxfp4_packed():
+            return super().dequant_model()
+
+        # Skipping base.py's dequant is only safe because the experts are the
+        # *only* quantized tensors. Verify that rather than assume it.
+        stray = [n for n in self.model_tensors
+                 if n.endswith(".weight_packed") and not self._MXFP4_EXPERT_RE.match(n)]
+        if stray:
+            raise NotImplementedError(
+                f"{len(stray)} MXFP4 tensor(s) outside the routed experts, e.g. {stray[0]!r}; "
+                "only the routed experts have a repack path"
+            )
+
+    def _mxfp4_expert_tensor(self, loaders: list[tuple[Callable[[], Tensor], Callable[[], Tensor]]]):
+        """
+        One stacked [n_expert, rows, cols] MXFP4 tensor, built lazily.
+
+        Laziness is not an optimization here, it is the difference between
+        working and not: gguf_writer holds every added tensor until the final
+        write, so materializing this eagerly (as the DeepSeek-V4 and NVFP4 paths
+        do) would keep all ~1.38 TB of experts resident. Deferring it means only
+        the tensor currently being written is in memory, one expert at a time.
+        """
+        # meta shapes, so this does not read any weights
+        rows, packed_cols = loaders[0][0]().shape
+        n_blocks = (packed_cols * 2) // 32
+        byte_shape = (len(loaders), rows, n_blocks * 17)
+
+        def load(fns: list[tuple[Callable[[], Tensor], Callable[[], Tensor]]]) -> np.ndarray:
+            out = np.empty(byte_shape, dtype=np.uint8)
+            for eid, (packed_fn, scale_fn) in enumerate(fns):
+                out[eid] = repack_mxfp4_blocks(
+                    LazyTorchTensor.to_eager(packed_fn()),
+                    LazyTorchTensor.to_eager(scale_fn()),
+                )
+            return out
+
+        # loaders goes through args rather than the closure so that `func` matches
+        # LazyBase's single-argument shape; _recurse_apply passes plain callables
+        # through untouched.
+        return gguf.LazyNumpyTensor(
+            meta=gguf.LazyNumpyTensor.meta_with_dtype_and_shape(np.uint8, byte_shape),
+            args=(loaders,),
+            func=load,
+        )
+
+    def _write_mxfp4_experts(self) -> None:
+        n_experts = self.hparams["num_experts"]
+
+        # (bid, wid) -> {expert id: (packed name, scale name)}
+        groups: dict[tuple[int, str], dict[int, tuple[str, str]]] = {}
+        for name in self.model_tensors:
+            m = self._MXFP4_EXPERT_RE.match(name)
+            if m is None:
+                continue
+            bid, eid, wid = int(m.group(1)), int(m.group(2)), m.group(3)
+            scale_name = name.removesuffix("_packed") + "_scale"
+            if scale_name not in self.model_tensors:
+                raise KeyError(f"missing {scale_name} for {name}")
+            groups.setdefault((bid, wid), {})[eid] = (name, scale_name)
+
+        consumed: list[str] = []
+        for (bid, wid), experts in sorted(groups.items()):
+            missing = [e for e in range(n_experts) if e not in experts]
+            if missing:
+                raise KeyError(
+                    f"layer {bid} {wid}: {len(missing)} of {n_experts} experts missing, "
+                    f"first is {missing[0]}"
+                )
+            if len(experts) != n_experts:
+                raise KeyError(f"layer {bid} {wid}: {len(experts)} experts, expected {n_experts}")
+
+            loaders = []
+            for eid in range(n_experts):
+                packed_name, scale_name = experts[eid]
+                loaders.append((self.model_tensors[packed_name], self.model_tensors[scale_name]))
+                consumed += [packed_name, scale_name]
+
+            data = self._mxfp4_expert_tensor(loaders)
+            new_name = self.format_tensor_name(self._MXFP4_PROJ[wid], bid)
+            shape = gguf.quant_shape_from_byte_shape(data.shape, gguf.GGMLQuantizationType.MXFP4)
+            logger.info(
+                f"{new_name}: repacked {n_experts} experts to MXFP4, "
+                f"shape = {{{', '.join(str(n) for n in reversed(shape))}}}"
+            )
+            self.gguf_writer.add_tensor(new_name, data, raw_dtype=gguf.GGMLQuantizationType.MXFP4)
+
+        for name in consumed:
+            del self.model_tensors[name]
+
+    def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
+        # Deliberately not a generator: base.py builds
+        # chain(generate_extra_tensors(), get_tensors()), so the tensors consumed
+        # here must be removed from model_tensors before get_tensors() starts.
+        if self._is_mxfp4_packed():
+            self._write_mxfp4_experts()
+        return ()
+
+    def get_tensors(self) -> Iterator[tuple[str, Tensor]]:
+        for name, data in super().get_tensors():
+            if name.startswith(("vision_tower.", "mm_projector.")):
+                continue  # text only
+            if name.startswith("language_model."):
+                name = name[len("language_model."):]
+            yield name, data
+
+    def set_gguf_parameters(self):
+        # MLA is served as MQA with a single large head, then decompressed
+        self.hparams["num_key_value_heads"] = 1
+
+        super().set_gguf_parameters()
+        self.gguf_writer.add_vocab_size(self.hparams["vocab_size"])
+
+        linear_attn_config = self.hparams["linear_attn_config"]
+
+        # layer types: n_head_kv == 0 marks a KDA (recurrent) layer.
+        # KimiLinearConfig.is_kda_layer uses (layer_idx + 1) in kda_layers, so
+        # the lists are 1-indexed - an off-by-one here silently produces garbage.
+        full_attn_layers = linear_attn_config["full_attn_layers"]
+        n_kv_heads = [
+            self.hparams["num_key_value_heads"] if (il + 1) in full_attn_layers else 0
+            for il in range(self.hparams["num_hidden_layers"])
+        ]
+        assert len(n_kv_heads) == self.hparams["num_hidden_layers"]
+        self.gguf_writer.add_head_count_kv(n_kv_heads)
+
+        # --- KDA ---
+        self.gguf_writer.add_ssm_conv_kernel(linear_attn_config["short_conv_kernel_size"])
+        self.gguf_writer.add_kda_head_dim(linear_attn_config["head_dim"])
+        if (lb := linear_attn_config.get("gate_lower_bound")) is not None:
+            self.gguf_writer.add_kda_gate_lower_bound(lb)
+
+        # --- MLA ---
+        if (q_lora_rank := self.hparams.get("q_lora_rank")) is not None:
+            self.gguf_writer.add_q_lora_rank(q_lora_rank)
+        kv_lora_rank = self.hparams["kv_lora_rank"]
+        self.gguf_writer.add_kv_lora_rank(kv_lora_rank)
+
+        qk_nope_head_dim = self.hparams["qk_nope_head_dim"]
+        qk_rope_head_dim = self.hparams["qk_rope_head_dim"]
+        v_head_dim = self.hparams["v_head_dim"]
+        # K3 is nope-only; qk_rope_head_dim still sizes the un-absorbed part of K
+        assert self.hparams.get("mla_use_nope"), "K3 MLA is expected to be nope-only"
+        self.gguf_writer.add_rope_dimension_count(qk_rope_head_dim)
+        self.gguf_writer.add_key_length(kv_lora_rank + qk_rope_head_dim)
+        self.gguf_writer.add_key_length_mla(qk_nope_head_dim + qk_rope_head_dim)
+        self.gguf_writer.add_value_length_mla(v_head_dim)
+
+        # --- MoE ---
+        self.gguf_writer.add_expert_feed_forward_length(self.hparams["moe_intermediate_size"])
+        self.gguf_writer.add_expert_shared_count(self.hparams["num_shared_experts"])
+        self.gguf_writer.add_leading_dense_block_count(self.hparams["first_k_dense_replace"])
+        self.gguf_writer.add_expert_weights_scale(self.hparams["routed_scaling_factor"])
+        self.gguf_writer.add_expert_weights_norm(self.hparams["moe_renormalize"])
+        assert self.hparams["moe_router_activation_func"] == "sigmoid"
+        self.gguf_writer.add_expert_gating_func(gguf.ExpertGatingFuncType.SIGMOID)
+        # latent MoE: routed experts live in a down-projected space
+        if (latent := self.hparams.get("routed_expert_hidden_size")) is not None:
+            self.gguf_writer.add_expert_latent_length(latent)
+
+        # --- situ activation ---
+        assert self.hparams["hidden_act"] == "situ", \
+            f"unexpected hidden_act {self.hparams['hidden_act']!r}"
+        self.gguf_writer.add_activation_situ_beta(self.hparams["activation_situ_beta"])
+        self.gguf_writer.add_activation_situ_linear_beta(self.hparams["activation_situ_linear_beta"])
+
+        # --- cross-layer attention residuals ---
+        self.gguf_writer.add_attn_res_block_size(self.hparams["attn_res_block_size"])
+
+    def prepare_tensors(self):
+        super().prepare_tensors()
+        if self._experts is not None:
+            leftover = [k for d in self._experts for k in d.keys()]
+            if leftover:
+                raise ValueError(f"Unprocessed experts: {leftover}")
+        if self._res_parts:
+            raise ValueError(f"Unpaired attention-residual tensors: {sorted(self._res_parts)}")
+        if self._is_mxfp4_packed():
+            # label the file for what it is; prepare_metadata runs after this
+            self._is_mxfp4 = True
+            self.ftype = gguf.LlamaFileType.MOSTLY_MXFP4_MOE
+
+    def _try_fuse_res(self, data_torch: Tensor, name: str, bid: int | None):
+        """
+        Pair <x>_res_norm.weight with <x>_res_proj.weight and emit their product.
+
+        proj is [1, n_embd]; norm is [n_embd]. _apply_attn_res only ever uses
+        norm.weight * proj.weight.squeeze(0), so one vector is enough.
+        Returns None if this is not a res tensor, [] if buffered pending its pair.
+        """
+        for prefix, (tensor_id, per_layer) in self._RES_FUSIONS.items():
+            for kind in ("norm", "proj"):
+                if not name.endswith(f"{prefix}_{kind}.weight"):
+                    continue
+                key = f"{prefix}.{bid}"
+                other = self._res_parts.pop(key, None)
+                if other is None:
+                    self._res_parts[key] = (kind, data_torch)
+                    return []
+                other_kind, other_data = other
+                assert other_kind != kind, f"duplicate {kind} for {key}"
+                norm = data_torch if kind == "norm" else other_data
+                proj = data_torch if kind == "proj" else other_data
+                fused = norm.float().flatten() * proj.float().flatten()
+                # ".weight" suffix matches the convention map_tensor_name applies
+                new_name = (self.format_tensor_name(tensor_id, bid) if per_layer
+                            else gguf.TENSOR_NAMES[tensor_id] + ".weight")
+                logger.info(f"fused {prefix}_norm * {prefix}_proj -> {new_name}")
+                return [(new_name, fused)]
+        return None
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # --- cross-layer attention residuals: fuse norm * proj ---
+        fused = self._try_fuse_res(data_torch, name, bid)
+        if fused is not None:
+            yield from fused
+            return
+
+        # --- KDA conv1d: HF [d_inner, 1, d_conv] -> ggml ne [d_conv, 1, d_inner, 1] ---
+        # GGUF reverses the numpy shape on write, so target numpy (1, d_inner, 1, d_conv).
+        # Both layouts have conv_step varying fastest, so this is a pure reshape.
+        if name.endswith((".q_conv1d.weight", ".k_conv1d.weight", ".v_conv1d.weight")):
+            if data_torch.ndim == 3:      # [d_inner, 1, d_conv]
+                d_inner, _, d_conv = data_torch.shape
+            elif data_torch.ndim == 2:    # [d_inner, d_conv]
+                d_inner, d_conv = data_torch.shape
+            else:
+                raise ValueError(f"unexpected conv1d rank {data_torch.ndim} for {name}")
+            data_torch = data_torch.reshape(1, d_inner, 1, d_conv)
+
+        # -exp(A_log) is folded here so the graph does not have to
+        if name.endswith(".A_log"):
+            n_head = self.hparams["num_attention_heads"]
+            data_torch = -torch.exp(data_torch.float()[:n_head])
+
+        # dt_bias -> the name SSM_DT's mapping expects
+        if name.endswith(".dt_bias"):
+            name = name.rpartition(".dt_bias")[0] + ".dt_proj.bias"
+
+        # --- g_proj is two different tensors sharing one HF name ---
+        # KDA layers: full-rank gate, [d_inner, n_embd]  (replaces g_a/g_b)
+        # MLA layers: output gate,    [n_head*v_head_dim, n_embd]
+        # Name-based mapping cannot tell them apart, so resolve by layer type.
+        if name.endswith(".self_attn.g_proj.weight"):
+            assert bid is not None
+            is_kda = (bid + 1) not in self.hparams["linear_attn_config"]["full_attn_layers"]
+            tensor_id = gguf.MODEL_TENSOR.SSM_G if is_kda else gguf.MODEL_TENSOR.ATTN_GATE
+            yield self.format_tensor_name(tensor_id, bid), data_torch
+            return
+
+        # --- routed experts: stack per-expert 2D weights into one 3D tensor ---
+        if ".block_sparse_moe.experts." in name:
+            n_experts = self.hparams["num_experts"]
+            assert bid is not None
+
+            if self._experts is None:
+                self._experts = [{} for _ in range(self.block_count)]
+            self._experts[bid][name] = data_torch
+
+            if len(self._experts[bid]) < n_experts * 3:
+                return
+
+            # w1: gate, w2: down, w3: up
+            for wid, tensor_id in (("w1", gguf.MODEL_TENSOR.FFN_GATE_EXP),
+                                   ("w2", gguf.MODEL_TENSOR.FFN_DOWN_EXP),
+                                   ("w3", gguf.MODEL_TENSOR.FFN_UP_EXP)):
+                datas = []
+                for xid in range(n_experts):
+                    ename = f"model.layers.{bid}.block_sparse_moe.experts.{xid}.{wid}.weight"
+                    datas.append(self._experts[bid].pop(ename))
+                stacked = torch.stack(datas, dim=0)
+                yield from super().modify_tensors(stacked, self.format_tensor_name(tensor_id, bid), bid)
+            return
+
+        # --- MLA absorption: split kv_b into k_b (transposed) and v_b ---
+        if name.endswith("kv_b_proj.weight"):
+            n_head_kv = self.hparams["num_key_value_heads"]
+            v_head_dim = self.hparams["v_head_dim"]
+            qk_nope_head_dim = self.hparams["qk_nope_head_dim"]
+            assert data_torch.shape[0] == n_head_kv * (v_head_dim + qk_nope_head_dim)
+            kv_b = data_torch.view(n_head_kv, v_head_dim + qk_nope_head_dim, data_torch.shape[-1])
+            k_b, v_b = torch.split(kv_b, [qk_nope_head_dim, v_head_dim], dim=1)
+            k_b = k_b.transpose(1, 2)
+            yield from super().modify_tensors(k_b, name.replace("kv_b_proj", "k_b_proj"), bid)
+            yield from super().modify_tensors(v_b, name.replace("kv_b_proj", "v_b_proj"), bid)
+            return
+
+        yield from super().modify_tensors(data_torch, name, bid)
diff --git a/conversion/kimivl.py b/conversion/kimivl.py
index 5ff3c39ca9c1..a5e278f1c090 100644
--- a/conversion/kimivl.py
+++ b/conversion/kimivl.py
@@ -168,3 +168,79 @@ def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iter
             name = name.replace("mm_projector.linear_", "mm_projector.proj.linear_", 1)
 
         yield from super().modify_tensors(data_torch, name, bid)
+
+
+@ModelBase.register("KimiK3ForConditionalGeneration")
+class KimiK3VisionModel(MmprojModel):
+    """Kimi-K3 MoonViT-3d vision tower (image path).
+
+    Structurally the Kimi-K2.5 tower with RMSNorm, no biases, a non-square fused QKV
+    (qkv_hidden_size 1536 vs vt_hidden_size 1024) and a post-norm patchmergerv2 projector.
+    Video is out of scope: for t == 1 the temporal pool and temporal position term vanish.
+    """
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        assert self.hparams_vision is not None, "Kimi-K3 requires vision_config in config.json"
+        self.merge_kernel_size = tuple(self.hparams_vision.get("merge_kernel_size", [2, 2]))
+        self.patch_size = self.hparams_vision.get("patch_size", 14)
+        pos_emb_h = self.hparams_vision.get("init_pos_emb_height", 64)
+        self.hparams_vision["image_size"] = pos_emb_h * self.patch_size
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        assert self.hparams_vision is not None
+        self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.KIMIK3)
+
+        # qkv width != n_embd, so the runtime cannot derive d_head
+        n_head = self.hparams_vision["vt_num_attention_heads"]
+        qkv_hidden = self.hparams_vision.get("qkv_hidden_size") or self.hparams_vision["vt_hidden_size"]
+        assert qkv_hidden % n_head == 0, f"qkv_hidden_size {qkv_hidden} not divisible by {n_head} heads"
+        self.gguf_writer.add_vision_head_dim(qkv_hidden // n_head)
+
+        self.gguf_writer.add_vision_use_gelu(True)  # activation_func is gelu_pytorch_tanh
+        self.gguf_writer.add_vision_attention_layernorm_eps(
+            self.hparams_vision.get("projector_ln_eps", 1e-5))
+        self.gguf_writer.add_vision_projector_scale_factor(self.merge_kernel_size[0])
+
+        in_patch_limit = self.preprocessor_config.get("media_proc_cfg", {}).get(
+            "in_patch_limit", self.preprocessor_config.get("in_patch_limit", 16384))
+        pixels_per_patch = self.patch_size ** 2
+        self.gguf_writer.add_vision_min_pixels(8 * pixels_per_patch)
+        self.gguf_writer.add_vision_max_pixels(in_patch_limit * pixels_per_patch)
+
+    @classmethod
+    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
+        name, _ = item
+        if not name.startswith(("vision_tower.", "mm_projector.")):
+            return None
+        return super().filter_tensors(item)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        assert self.hparams_vision is not None
+        n_head = self.hparams_vision["vt_num_attention_heads"]
+
+        if "wqkv" in name and "weight" in name:
+            # de-interleave Q/K so the runtime can use build_rope_2d(interleave_freq=false)
+            out_dim = data_torch.shape[0]
+            qkv_dim = out_dim // 3
+            head_dim = qkv_dim // n_head
+            wq, wk, wv = (data_torch[:qkv_dim], data_torch[qkv_dim:2 * qkv_dim], data_torch[2 * qkv_dim:])
+
+            def deinterleave(w: Tensor) -> Tensor:
+                return (w.reshape(n_head, head_dim // 4, 2, 2, w.shape[-1])
+                         .permute(0, 2, 1, 3, 4)
+                         .reshape(w.shape[0], w.shape[-1]))
+
+            data_torch = torch.cat([deinterleave(wq), deinterleave(wk), wv], dim=0)
+
+        if "pos_emb.weight" in name:
+            # kept 3D: the runtime reads grid extents from ne[1]/ne[2]
+            pass
+
+        if "mm_projector.proj.0." in name:
+            name = name.replace(".proj.0.", ".proj.linear_1.")
+        elif "mm_projector.proj.2." in name:
+            name = name.replace(".proj.2.", ".proj.linear_2.")
+
+        yield from super().modify_tensors(data_torch, name, bid)
diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py
index 643063e662e0..5479e5bb12a1 100644
--- a/gguf-py/gguf/constants.py
+++ b/gguf-py/gguf/constants.py
@@ -123,6 +123,7 @@ class LLM:
         EXPERT_WEIGHTS_NORM               = "{arch}.expert_weights_norm"
         EXPERT_GATING_FUNC                = "{arch}.expert_gating_func"
         EXPERT_GROUP_SCALE                = "{arch}.expert_group_scale"
+        EXPERT_LATENT_LENGTH              = "{arch}.expert_latent_length"
         EXPERTS_PER_GROUP                 = "{arch}.experts_per_group"
         MOE_EVERY_N_LAYERS                = "{arch}.moe_every_n_layers"
         MOE_LATENT_SIZE                   = "{arch}.moe_latent_size"
@@ -230,6 +231,13 @@ class Rope:
         SCALING_YARN_BETA_FAST    = "{arch}.rope.scaling.yarn_beta_fast"
         SCALING_YARN_BETA_SLOW    = "{arch}.rope.scaling.yarn_beta_slow"
 
+    class Activation:
+        SITU_BETA        = "{arch}.activation.situ_beta"
+        SITU_LINEAR_BETA = "{arch}.activation.situ_linear_beta"
+
+    class AttnRes:
+        BLOCK_SIZE = "{arch}.attn_res.block_size"
+
     class Split:
         LLM_KV_SPLIT_NO            = "split.no"
         LLM_KV_SPLIT_COUNT         = "split.count"
@@ -244,7 +252,8 @@ class SSM:
         DT_B_C_RMS     = "{arch}.ssm.dt_b_c_rms"
 
     class KDA:
-        HEAD_DIM = "{arch}.kda.head_dim"
+        HEAD_DIM         = "{arch}.kda.head_dim"
+        GATE_LOWER_BOUND = "{arch}.kda.gate_lower_bound"
 
     class WKV:
         HEAD_SIZE = "{arch}.wkv.head_size"
@@ -553,6 +562,7 @@ class MODEL_ARCH(IntEnum):
     LLAMA_EMBED      = auto()
     MAINCODER        = auto()
     KIMI_LINEAR      = auto()
+    KIMI_K3          = auto()
     TALKIE           = auto()
     MELLUM           = auto()
     NANBEIGE         = auto()
@@ -670,6 +680,13 @@ class MODEL_TENSOR(IntEnum):
     SSM_BETA             = auto() # Kimi Linear qwen3.5
     SSM_G_A              = auto() # Kimi Linear
     SSM_G_B              = auto() # Kimi Linear
+    SSM_G                = auto() # Kimi K3 (full-rank KDA gate, replaces SSM_G_A/SSM_G_B)
+    ATTN_RES_SCORE       = auto() # Kimi K3 (fused res_norm * res_proj, pre-attention)
+    FFN_RES_SCORE        = auto() # Kimi K3 (fused res_norm * res_proj, pre-FFN)
+    OUTPUT_RES_SCORE     = auto() # Kimi K3 (fused res_norm * res_proj, final)
+    FFN_ROUTED_DOWN      = auto() # Kimi K3 (latent MoE: hidden -> latent)
+    FFN_ROUTED_UP        = auto() # Kimi K3 (latent MoE: latent -> hidden)
+    FFN_ROUTED_NORM      = auto() # Kimi K3 (latent MoE: norm on expert output)
     TIME_MIX_W0          = auto()
     TIME_MIX_W1          = auto()
     TIME_MIX_W2          = auto()
@@ -1174,6 +1191,7 @@ class MODEL_TENSOR(IntEnum):
     MODEL_ARCH.LLAMA_EMBED:      "llama-embed",
     MODEL_ARCH.MAINCODER:        "maincoder",
     MODEL_ARCH.KIMI_LINEAR:      "kimi-linear",
+    MODEL_ARCH.KIMI_K3:          "kimi-k3",
     MODEL_ARCH.TALKIE:           "talkie",
     MODEL_ARCH.MELLUM:           "mellum",
     MODEL_ARCH.NANBEIGE:         "nanbeige",
@@ -1289,6 +1307,13 @@ class MODEL_TENSOR(IntEnum):
     MODEL_TENSOR.SSM_BETA:                  "blk.{bid}.ssm_beta",             # Kimi Linear qwen3.5
     MODEL_TENSOR.SSM_G_A:                   "blk.{bid}.ssm_g_a",              # Kimi Linear
     MODEL_TENSOR.SSM_G_B:                   "blk.{bid}.ssm_g_b",              # Kimi Linear
+    MODEL_TENSOR.SSM_G:                     "blk.{bid}.ssm_g",                # Kimi K3
+    MODEL_TENSOR.ATTN_RES_SCORE:            "blk.{bid}.attn_res_score",       # Kimi K3
+    MODEL_TENSOR.FFN_RES_SCORE:             "blk.{bid}.ffn_res_score",        # Kimi K3
+    MODEL_TENSOR.OUTPUT_RES_SCORE:          "output_res_score",               # Kimi K3
+    MODEL_TENSOR.FFN_ROUTED_DOWN:           "blk.{bid}.ffn_routed_down",      # Kimi K3
+    MODEL_TENSOR.FFN_ROUTED_UP:             "blk.{bid}.ffn_routed_up",        # Kimi K3
+    MODEL_TENSOR.FFN_ROUTED_NORM:           "blk.{bid}.ffn_routed_norm",      # Kimi K3
     MODEL_TENSOR.TIME_MIX_W0:               "blk.{bid}.time_mix_w0",
     MODEL_TENSOR.TIME_MIX_W1:               "blk.{bid}.time_mix_w1",
     MODEL_TENSOR.TIME_MIX_W2:               "blk.{bid}.time_mix_w2",
@@ -4602,6 +4627,56 @@ class MODEL_TENSOR(IntEnum):
         MODEL_TENSOR.FFN_DOWN_SHEXP,
         MODEL_TENSOR.FFN_UP_SHEXP,
     ],
+    MODEL_ARCH.KIMI_K3: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.OUTPUT_RES_SCORE,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_RES_SCORE,
+        MODEL_TENSOR.FFN_RES_SCORE,
+        # MLA (full-attention layers)
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.ATTN_GATE,
+        MODEL_TENSOR.ATTN_Q_A,
+        MODEL_TENSOR.ATTN_Q_B,
+        MODEL_TENSOR.ATTN_KV_A_MQA,
+        MODEL_TENSOR.ATTN_KV_B,
+        MODEL_TENSOR.ATTN_K_B,
+        MODEL_TENSOR.ATTN_V_B,
+        MODEL_TENSOR.ATTN_Q_A_NORM,
+        MODEL_TENSOR.ATTN_KV_A_NORM,
+        # KDA (linear-attention layers)
+        MODEL_TENSOR.SSM_CONV1D_Q,
+        MODEL_TENSOR.SSM_CONV1D_K,
+        MODEL_TENSOR.SSM_CONV1D_V,
+        MODEL_TENSOR.SSM_F_A,
+        MODEL_TENSOR.SSM_F_B,
+        MODEL_TENSOR.SSM_BETA,
+        MODEL_TENSOR.SSM_A,
+        MODEL_TENSOR.SSM_G,
+        MODEL_TENSOR.SSM_DT,
+        MODEL_TENSOR.SSM_NORM,
+        # FFN
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+        MODEL_TENSOR.FFN_GATE_INP,
+        MODEL_TENSOR.FFN_EXP_PROBS_B,
+        MODEL_TENSOR.FFN_GATE_EXP,
+        MODEL_TENSOR.FFN_DOWN_EXP,
+        MODEL_TENSOR.FFN_UP_EXP,
+        MODEL_TENSOR.FFN_GATE_SHEXP,
+        MODEL_TENSOR.FFN_DOWN_SHEXP,
+        MODEL_TENSOR.FFN_UP_SHEXP,
+        MODEL_TENSOR.FFN_ROUTED_DOWN,
+        MODEL_TENSOR.FFN_ROUTED_UP,
+        MODEL_TENSOR.FFN_ROUTED_NORM,
+    ],
     MODEL_ARCH.TALKIE: [
         MODEL_TENSOR.TOKEN_EMBD,
         MODEL_TENSOR.OUTPUT,
@@ -4912,6 +4987,7 @@ class VisionProjectorType:
     KIMIVL = "kimivl"
     PADDLEOCR = "paddleocr"
     KIMIK25 = "kimik25"
+    KIMIK3 = "kimik3"
     LIGHTONOCR = "lightonocr"
     COGVLM = "cogvlm"
     JANUS_PRO = "janus_pro"
diff --git a/gguf-py/gguf/gguf_writer.py b/gguf-py/gguf/gguf_writer.py
index 3aa4f049f2bb..a7f4a9eb3ac1 100644
--- a/gguf-py/gguf/gguf_writer.py
+++ b/gguf-py/gguf/gguf_writer.py
@@ -1084,6 +1084,21 @@ def add_ssm_group_count(self, value: int) -> None:
     def add_ssm_dt_b_c_rms(self, value: bool) -> None:
         self.add_bool(Keys.SSM.DT_B_C_RMS.format(arch=self.arch), value)
 
+    def add_kda_gate_lower_bound(self, value: float) -> None:
+        self.add_float32(Keys.KDA.GATE_LOWER_BOUND.format(arch=self.arch), value)
+
+    def add_expert_latent_length(self, value: int) -> None:
+        self.add_uint32(Keys.LLM.EXPERT_LATENT_LENGTH.format(arch=self.arch), value)
+
+    def add_activation_situ_beta(self, value: float) -> None:
+        self.add_float32(Keys.Activation.SITU_BETA.format(arch=self.arch), value)
+
+    def add_activation_situ_linear_beta(self, value: float) -> None:
+        self.add_float32(Keys.Activation.SITU_LINEAR_BETA.format(arch=self.arch), value)
+
+    def add_attn_res_block_size(self, value: int) -> None:
+        self.add_uint32(Keys.AttnRes.BLOCK_SIZE.format(arch=self.arch), value)
+
     def add_kda_head_dim(self, value: int) -> None:
         self.add_uint32(Keys.KDA.HEAD_DIM.format(arch=self.arch), value)
 
diff --git a/gguf-py/gguf/tensor_mapping.py b/gguf-py/gguf/tensor_mapping.py
index 5a0175a56c8e..48656e1f6437 100644
--- a/gguf-py/gguf/tensor_mapping.py
+++ b/gguf-py/gguf/tensor_mapping.py
@@ -912,6 +912,19 @@ class TensorNameMap:
             "model.layers.{bid}.linear_attn.in_proj_b",  # qwen3.5
             "model.layers.{bid}.self_attn.b_proj",       # Kimi Linear
         ),
+        # Kimi K3 latent MoE: routed experts operate in a down-projected space
+        MODEL_TENSOR.FFN_ROUTED_DOWN: (
+            "model.layers.{bid}.block_sparse_moe.routed_expert_down_proj",
+        ),
+
+        MODEL_TENSOR.FFN_ROUTED_UP: (
+            "model.layers.{bid}.block_sparse_moe.routed_expert_up_proj",
+        ),
+
+        MODEL_TENSOR.FFN_ROUTED_NORM: (
+            "model.layers.{bid}.block_sparse_moe.routed_expert_norm",
+        ),
+
         MODEL_TENSOR.SSM_G_A: (
             "model.layers.{bid}.self_attn.g_a_proj",
         ),
diff --git a/models/templates/Kimi-K3.jinja b/models/templates/Kimi-K3.jinja
new file mode 100644
index 000000000000..67e51a6af4b3
--- /dev/null
+++ b/models/templates/Kimi-K3.jinja
@@ -0,0 +1,325 @@
+{%- macro escape_attr(value) -%}
+{{- value|string|replace('&', '&amp;')|replace('"', '&quot;') -}}
+{%- endmacro -%}
+
+{%- macro open_tag(tag, attrs=[]) -%}
+{{- '<|open|>' + tag -}}
+{%- for attr in attrs -%}
+{{- ' ' + attr[0] + '="' -}}{{- escape_attr(attr[1]) -}}{{- '"' -}}
+{%- endfor -%}
+{{- '<|sep|>' -}}
+{%- endmacro -%}
+
+{%- macro close_tag(tag) -%}
+{{- '<|close|>' + tag + '<|sep|>' -}}
+{%- endmacro -%}
+
+{%- macro next_image(state) -%}
+{%- if image_prompts is defined and image_prompts is not none -%}
+    {%- if state.image_index >= image_prompts|length -%}
+        {{- raise_exception('More image placeholders than image prompts.') -}}
+    {%- endif -%}
+    {{- image_prompts[state.image_index] -}}
+    {%- set state.image_index = state.image_index + 1 -%}
+{%- else -%}
+    {{- '<|kimi_image_placeholder|>' -}}
+{%- endif -%}
+{%- endmacro -%}
+
+{%- macro render_text(text, state) -%}
+{%- set text = text|string -%}
+{%- if image_prompts is defined and image_prompts is not none and '<|kimi_image_placeholder|>' in text -%}
+    {%- set parts = text.split('<|kimi_image_placeholder|>') -%}
+    {%- for part in parts -%}
+        {{- part -}}
+        {%- if not loop.last -%}{{- next_image(state) -}}{%- endif -%}
+    {%- endfor -%}
+{%- else -%}
+    {{- text -}}
+{%- endif -%}
+{%- endmacro -%}
+
+{%- macro render_content(content, state) -%}
+{%- if content is string -%}
+    {{- render_text(content, state) -}}
+{%- elif content is not none and content is defined -%}
+    {%- for part in content -%}
+        {%- if part.type in ['image', 'image_url'] -%}
+            {{- next_image(state) -}}
+        {%- else -%}
+            {{- render_text(part.text, state) -}}
+        {%- endif -%}
+    {%- endfor -%}
+{%- endif -%}
+{%- endmacro -%}
+
+{%- macro internal_system_message(message_type, body) -%}
+{{- open_tag('message', [('role', 'system'), ('type', message_type)]) -}}
+{{- body|trim -}}
+{{- close_tag('message') -}}
+{{- '<|end_of_msg|>' -}}
+{%- endmacro -%}
+
+{%- macro json_sorted(value) -%}
+{#- Minja は tojson(sort_keys=true) を実装していない。K3 参照実装は
+    deep_sort_dict() の後に compact JSON 化するため、dictsort と再帰マクロで
+    同じバイト列を作る。配列の順序は保持し、mapping の各階層だけソートする。 -#}
+{%- if value is mapping -%}
+{{- '{' -}}
+{%- for key, item in value|dictsort -%}
+{%- if not loop.first -%}{{- ',' -}}{%- endif -%}
+{{- key|tojson(ensure_ascii=false) -}}{{- ':' -}}{{- json_sorted(item) -}}
+{%- endfor -%}
+{{- '}' -}}
+{%- elif value is string or value is number or value is boolean or value is none -%}
+{{- value|tojson(ensure_ascii=false) -}}
+{%- else -%}
+{{- '[' -}}
+{%- for item in value -%}
+{%- if not loop.first -%}{{- ',' -}}{%- endif -%}
+{{- json_sorted(item) -}}
+{%- endfor -%}
+{{- ']' -}}
+{%- endif -%}
+{%- endmacro -%}
+
+{%- macro render_tool_declare(tool_list, dynamic=false) -%}
+{{- open_tag('message', [('role', 'system'), ('type', 'tool-declare')]) -}}
+{%- if dynamic -%}
+{{- '## New Tools Available\nThe system dynamically extends the toolset via lazy-loading.\nYou have access to all existing and extended tools.\nHere are the specs for the extended tools.\n\n```json\n' -}}
+{%- else -%}
+{{- '# Tools\nHere are the available tools, described in JSONSchema.\n\n```json\n' -}}
+{%- endif -%}
+{{- json_sorted(tool_list) -}}
+{{- '\n```' -}}
+{{- close_tag('message') -}}
+{{- '<|end_of_msg|>' -}}
+{%- endmacro -%}
+
+{%- macro xtml_type(value) -%}
+{%- if value is boolean -%}boolean
+{%- elif value is none -%}null
+{%- elif value is number -%}number
+{%- elif value is string -%}string
+{%- elif value is mapping -%}object
+{%- else -%}array
+{%- endif -%}
+{%- endmacro -%}
+
+{%- macro xtml_value(value) -%}
+{%- if value is string -%}
+{{- value -}}
+{%- else -%}
+{{- value|tojson(ensure_ascii=false) -}}
+{%- endif -%}
+{%- endmacro -%}
+
+{%- macro render_assistant(message, state) -%}
+{%- if thinking -%}
+    {%- set reasoning_content = message.get('reasoning_content') or message.get('reasoning') -%}
+    {{- open_tag('think') -}}
+    {%- if reasoning_content is not none and reasoning_content|string|trim -%}
+        {{- render_text(reasoning_content, state) -}}
+    {%- endif -%}
+    {{- close_tag('think') -}}
+{%- endif -%}
+{{- open_tag('response') -}}
+{{- render_content(message.get('content'), state) -}}
+{{- close_tag('response') -}}
+{%- set tool_calls = message.get('tool_calls') -%}
+{%- if tool_calls -%}
+    {{- open_tag('tools') -}}
+    {%- for tool_call in tool_calls -%}
+        {%- if tool_call is not mapping -%}
+            {{- raise_exception('Kimi K3 tool calls must be mappings.') -}}
+        {%- endif -%}
+        {%- set fn = tool_call.function if tool_call.function is defined and tool_call.function is mapping else tool_call -%}
+        {%- if fn.get('name') is none -%}
+            {{- raise_exception('Kimi K3 tool calls require a function name.') -}}
+        {%- endif -%}
+        {{- open_tag('call', [('tool', fn.name), ('index', loop.index)]) -}}
+        {%- set arguments = fn.get('arguments', {}) -%}
+        {%- set json_block = fn.get('_xtml_json_block') -%}
+        {%- if json_block is not none -%}
+            {{- open_tag('json', [('type', 'object')]) -}}
+            {{- render_text(json_block, state) -}}
+            {{- close_tag('json') -}}
+        {%- elif arguments is mapping -%}
+            {%- for key, value in arguments.items() -%}
+                {{- open_tag('argument', [('key', key), ('type', xtml_type(value))]) -}}
+                {{- render_text(xtml_value(value), state) -}}
+                {{- close_tag('argument') -}}
+            {%- endfor -%}
+        {%- elif arguments is string and arguments|trim -%}
+            {{- open_tag('json', [('type', 'object')]) -}}
+            {{- render_text(arguments, state) -}}
+            {{- close_tag('json') -}}
+        {%- elif arguments is not none and arguments is not string -%}
+            {{- raise_exception('Kimi K3 tool call arguments must be a mapping or a JSON object string.') -}}
+        {%- endif -%}
+        {{- close_tag('call') -}}
+    {%- endfor -%}
+    {{- close_tag('tools') -}}
+{%- endif -%}
+{%- endmacro -%}
+
+{%- macro render_tool_message(message, state, resolved_name=none) -%}
+{%- set state.tool_index = state.tool_index + 1 -%}
+{%- if resolved_name is not none -%}
+    {%- set tool_name = resolved_name -%}
+{%- elif 'tool' in message -%}
+    {%- set tool_name = message.get('tool') -%}
+{%- else -%}
+    {%- set tool_name = message.get('name') -%}
+{%- endif -%}
+{%- if tool_name is none and state.tool_calls is not none and state.tool_index <= state.tool_calls|length -%}
+    {%- set fallback_call = state.tool_calls[state.tool_index - 1] -%}
+    {%- set fallback_fn = fallback_call.function if fallback_call.function is defined and fallback_call.function is mapping else fallback_call -%}
+    {%- set tool_name = fallback_fn.name -%}
+{%- endif -%}
+{%- if tool_name is none -%}
+    {{- raise_exception('Kimi K3 tool messages need a resolvable tool name: carry `tool`/`name`, or match a preceding assistant tool_call by order.') -}}
+{%- endif -%}
+{{- open_tag('message', [('role', 'tool'), ('tool', tool_name), ('index', state.tool_index)]) -}}
+{{- render_content(message.get('content'), state) -}}
+{{- close_tag('message') -}}
+{{- '<|end_of_msg|>' -}}
+{%- endmacro -%}
+
+{%- if thinking is undefined -%}
+    {%- set thinking = true -%}
+{%- endif -%}
+{%- if thinking_effort is undefined -%}
+    {%- set thinking_effort = 'max' -%}
+{%- endif -%}
+{%- if thinking and thinking_effort is not none and thinking_effort not in ['low', 'high', 'max'] -%}
+    {{- raise_exception('Unsupported thinking_effort=' + thinking_effort|string + '; supported values are low, high, and max.') -}}
+{%- endif -%}
+
+{%- set state = namespace(image_index=0, tool_calls=none, tool_index=0, response_schema=none) -%}
+
+{%- if tools is defined and tools -%}
+    {{- render_tool_declare(tools) -}}
+{%- endif -%}
+
+{%- if thinking and thinking_effort in ['low', 'high', 'max'] -%}
+    {{- internal_system_message(
+        'thinking-effort',
+        '`thinking_effort` guides on how much to think in your thinking channel (not including the response channel), supported values include `low`, `medium`, `high`, and `max`.\nNow the system is invoked with `thinking_effort=' + thinking_effort|string + '`.'
+    ) -}}
+{%- endif -%}
+
+{%- for message in messages -%}
+    {%- if message is mapping -%}
+        {%- if 'role' not in message -%}
+            {{- raise_exception('Kimi K3 messages require a role.') -}}
+        {%- elif message.role == 'user' -%}
+            {%- set attrs = [('role', 'user')] -%}
+            {%- if message.get('name') -%}{%- set attrs = attrs + [('name', message.name)] -%}{%- endif -%}
+            {{- open_tag('message', attrs) -}}
+            {{- render_content(message.get('content'), state) -}}
+            {{- close_tag('message') -}}
+            {{- '<|end_of_msg|>' -}}
+        {%- elif message.role == 'system' and message.get('tools') -%}
+            {{- render_tool_declare(message.tools, dynamic=true) -}}
+        {%- elif message.role == 'system' -%}
+            {%- set attrs = [('role', 'system')] -%}
+            {%- if message.get('name') -%}{%- set attrs = attrs + [('name', message.name)] -%}{%- endif -%}
+            {{- open_tag('message', attrs) -}}
+            {{- render_content(message.get('content'), state) -}}
+            {{- close_tag('message') -}}
+            {{- '<|end_of_msg|>' -}}
+        {%- elif message.role == 'assistant' -%}
+            {%- set state.tool_calls = message.get('tool_calls') -%}
+            {%- set state.tool_index = 0 -%}
+            {%- set attrs = [('role', 'assistant')] -%}
+            {%- if message.get('name') -%}{%- set attrs = attrs + [('name', message.name)] -%}{%- endif -%}
+            {{- open_tag('message', attrs) -}}
+            {{- render_assistant(message, state) -}}
+            {{- close_tag('message') -}}
+            {{- '<|end_of_msg|>' -}}
+        {%- elif message.role == 'tool' and (loop.first or messages[loop.index0 - 1].role != 'tool') -%}
+            {%- set run = namespace(tool_messages=[], resolved_count=0) -%}
+            {%- for candidate in messages[loop.index0:] -%}
+                {%- if candidate is not mapping or candidate.role != 'tool' -%}{%- break -%}{%- endif -%}
+                {%- set run.tool_messages = run.tool_messages + [candidate] -%}
+                {%- set call_id = candidate.get('tool_call_id', candidate.get('id')) -%}
+                {%- set match = namespace(found=false) -%}
+                {%- if call_id is not none and state.tool_calls is not none -%}
+                    {%- for tool_call in state.tool_calls -%}
+                        {%- if not match.found and tool_call is mapping and tool_call.get('id') is not none and tool_call.get('id')|string == call_id|string -%}
+                            {%- set match.found = true -%}
+                        {%- endif -%}
+                    {%- endfor -%}
+                {%- endif -%}
+                {%- if match.found -%}{%- set run.resolved_count = run.resolved_count + 1 -%}{%- endif -%}
+            {%- endfor -%}
+            {%- if run.tool_messages|length > 0 and run.resolved_count == run.tool_messages|length -%}
+                {%- set emitted = namespace(ids=[]) -%}
+                {%- for tool_call in state.tool_calls -%}
+                    {%- if tool_call is mapping and tool_call.get('id') is not none and tool_call.get('id')|string not in emitted.ids -%}
+                        {%- set emitted.ids = emitted.ids + [tool_call.get('id')|string] -%}
+                        {%- set fn = tool_call.function if tool_call.function is defined and tool_call.function is mapping else tool_call -%}
+                        {%- for tool_message in run.tool_messages -%}
+                            {%- set result_id = tool_message.get('tool_call_id', tool_message.get('id')) -%}
+                            {%- if result_id is not none and result_id|string == tool_call.get('id')|string -%}
+                                {{- render_tool_message(tool_message, state, fn.get('name')) -}}
+                            {%- endif -%}
+                        {%- endfor -%}
+                    {%- endif -%}
+                {%- endfor -%}
+            {%- else -%}
+                {%- for tool_message in run.tool_messages -%}
+                    {{- render_tool_message(tool_message, state) -}}
+                {%- endfor -%}
+            {%- endif -%}
+        {%- endif -%}
+    {%- endif -%}
+{%- endfor -%}
+
+{%- if tool_choice is defined and tool_choice == 'required' -%}
+    {{- internal_system_message('tool-choice', 'The system is invoked with `tool_choice=required`.\nYou MUST call tools in the next message.') -}}
+{%- elif tool_choice is defined and tool_choice == 'none' -%}
+    {{- internal_system_message('tool-choice', 'The system is invoked with `tool_choice=none`.\nYou MUST NOT call any tools in the next message.') -}}
+{%- endif -%}
+
+{%- if response_schema is defined -%}
+    {%- set state.response_schema = response_schema -%}
+{%- elif response_format is defined and response_format is mapping and response_format.get('json_schema') is not none -%}
+    {%- set schema_wrapper = response_format.get('json_schema') -%}
+    {%- if schema_wrapper is mapping and 'schema' in schema_wrapper -%}
+        {%- set state.response_schema = schema_wrapper.get('schema') -%}
+    {%- elif schema_wrapper is mapping and 'json_schema' in schema_wrapper -%}
+        {%- set state.response_schema = schema_wrapper.get('json_schema') -%}
+    {%- else -%}
+        {%- set state.response_schema = schema_wrapper -%}
+    {%- endif -%}
+{%- endif -%}
+
+{%- set response_format_type = none -%}
+{%- if response_format is defined and response_format is mapping -%}
+    {%- set response_format_type = response_format.get('type') -%}
+{%- elif response_format is defined -%}
+    {%- set response_format_type = response_format -%}
+{%- endif -%}
+{%- if response_format_type == 'json_object' -%}
+    {{- internal_system_message(
+        'response-format',
+        'The system is invoked with `response_format=json_object`.\nYour response must be raw JSON data without markdown code blocks (```json) or any additional formatting.'
+    ) -}}
+{%- elif response_format_type == 'json_schema' -%}
+    {{- internal_system_message(
+        'response-format',
+        'The system is invoked with `response_format=json_schema`.\nYour response must be raw JSON data without markdown code blocks (```json) or any additional formatting.\nThe JSON data must match the following schema:\n```json\n' + json_sorted(state.response_schema) + '\n```'
+    ) -}}
+{%- endif -%}
+
+{%- if add_generation_prompt -%}
+    {{- open_tag('message', [('role', 'assistant')]) -}}
+    {{- open_tag('think' if thinking else 'response') -}}
+{%- endif -%}
+
+{%- if image_prompts is defined and image_prompts is not none and state.image_index != image_prompts|length -%}
+    {{- raise_exception('image prompt count ' + image_prompts|length|string + ' != consumed placeholder count ' + state.image_index|string) -}}
+{%- endif -%}
+
diff --git a/src/llama-arch.cpp b/src/llama-arch.cpp
index 99fad9738a06..8aeb2bdcd4ee 100644
--- a/src/llama-arch.cpp
+++ b/src/llama-arch.cpp
@@ -141,6 +141,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_LLAMA_EMBED,      "llama-embed"      },
     { LLM_ARCH_MAINCODER,        "maincoder"        },
     { LLM_ARCH_KIMI_LINEAR,      "kimi-linear"      },
+    { LLM_ARCH_KIMI_K3,          "kimi-k3"          },
     { LLM_ARCH_TALKIE,           "talkie"           },
     { LLM_ARCH_MELLUM,           "mellum"           },
     { LLM_ARCH_NANBEIGE,         "nanbeige"         },
@@ -183,6 +184,9 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_FEATURES_LENGTH,                   "%s.features_length"                   },
     { LLM_KV_BLOCK_COUNT,                       "%s.block_count"                       },
     { LLM_KV_LEADING_DENSE_BLOCK_COUNT,         "%s.leading_dense_block_count"         },
+    { LLM_KV_ATTN_RES_BLOCK_SIZE,               "%s.attn_res.block_size"               },
+    { LLM_KV_ACTIVATION_SITU_BETA,              "%s.activation.situ_beta"              },
+    { LLM_KV_ACTIVATION_SITU_LINEAR_BETA,       "%s.activation.situ_linear_beta"       },
     { LLM_KV_FEED_FORWARD_LENGTH,               "%s.feed_forward_length"               },
     { LLM_KV_EXPERT_FEED_FORWARD_LENGTH,        "%s.expert_feed_forward_length"        },
     { LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, "%s.expert_shared_feed_forward_length" },
@@ -198,6 +202,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_EXPERT_GROUP_USED_COUNT,           "%s.expert_group_used_count"           },
     { LLM_KV_EXPERT_WEIGHTS_SCALE,              "%s.expert_weights_scale"              },
     { LLM_KV_EXPERT_WEIGHTS_NORM,               "%s.expert_weights_norm"               },
+    { LLM_KV_EXPERT_LATENT_LENGTH,              "%s.expert_latent_length"              },
     { LLM_KV_EXPERT_GATING_FUNC,                "%s.expert_gating_func"                },
     { LLM_KV_EXPERT_GROUP_SCALE,                "%s.expert_group_scale"                },
     { LLM_KV_EXPERTS_PER_GROUP,                 "%s.experts_per_group"                 },
@@ -304,6 +309,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_SSM_DT_B_C_RMS,     "%s.ssm.dt_b_c_rms"     },
 
     { LLM_KV_KDA_HEAD_DIM, "%s.kda.head_dim" },
+    { LLM_KV_KDA_GATE_LOWER_BOUND, "%s.kda.gate_lower_bound" },
 
     { LLM_KV_WKV_HEAD_SIZE, "%s.wkv.head_size" },
 
@@ -466,6 +472,13 @@ static const std::map<llm_tensor, const char *> LLM_TENSOR_NAMES = {
     { LLM_TENSOR_SSM_F_B,                                "blk.%d.ssm_f_b" },
     { LLM_TENSOR_SSM_BETA,                               "blk.%d.ssm_beta" },
     { LLM_TENSOR_SSM_G_A,                                "blk.%d.ssm_g_a" },
+    { LLM_TENSOR_SSM_G,                                  "blk.%d.ssm_g" },
+    { LLM_TENSOR_ATTN_RES_SCORE,                         "blk.%d.attn_res_score" },
+    { LLM_TENSOR_FFN_RES_SCORE,                          "blk.%d.ffn_res_score" },
+    { LLM_TENSOR_OUTPUT_RES_SCORE,                       "output_res_score" },
+    { LLM_TENSOR_FFN_ROUTED_DOWN,                        "blk.%d.ffn_routed_down" },
+    { LLM_TENSOR_FFN_ROUTED_UP,                          "blk.%d.ffn_routed_up" },
+    { LLM_TENSOR_FFN_ROUTED_NORM,                        "blk.%d.ffn_routed_norm" },
     { LLM_TENSOR_SSM_G_B,                                "blk.%d.ssm_g_b" },
     { LLM_TENSOR_SSM_NORM,                               "blk.%d.ssm_norm" },
     { LLM_TENSOR_ATTN_Q_A_NORM,                          "blk.%d.attn_q_a_norm" },
@@ -769,6 +782,13 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
     {LLM_TENSOR_SSM_F_B,                    {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_SSM_BETA,                   {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_SSM_G_A,                    {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_SSM_G,                      {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_ATTN_RES_SCORE,             {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
+    {LLM_TENSOR_FFN_RES_SCORE,              {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
+    {LLM_TENSOR_OUTPUT_RES_SCORE,           {LLM_TENSOR_LAYER_OUTPUT,    GGML_OP_MUL}},
+    {LLM_TENSOR_FFN_ROUTED_DOWN,            {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_FFN_ROUTED_UP,              {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_FFN_ROUTED_NORM,            {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
     {LLM_TENSOR_SSM_G_B,                    {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_TIME_MIX_LERP_X,            {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
     {LLM_TENSOR_TIME_MIX_LN,                {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
@@ -999,6 +1019,7 @@ bool llm_arch_is_hybrid(const llm_arch & arch) {
         case LLM_ARCH_NEMOTRON_H_MOE:
         case LLM_ARCH_QWEN3NEXT:
         case LLM_ARCH_KIMI_LINEAR:
+        case LLM_ARCH_KIMI_K3:
         case LLM_ARCH_QWEN35:
         case LLM_ARCH_QWEN35MOE:
         case LLM_ARCH_INKLING:
@@ -1058,6 +1079,7 @@ bool llm_arch_supports_sm_tensor(const llm_arch & arch) {
         case LLM_ARCH_MINIMAX_M3:
         case LLM_ARCH_MISTRAL4:
         case LLM_ARCH_KIMI_LINEAR:
+        case LLM_ARCH_KIMI_K3:
         case LLM_ARCH_INKLING:
             return false;
         default:
diff --git a/src/llama-arch.h b/src/llama-arch.h
index 6a283eb6b5a1..1eb2f5bcbb94 100644
--- a/src/llama-arch.h
+++ b/src/llama-arch.h
@@ -143,6 +143,7 @@ enum llm_arch {
     LLM_ARCH_LLAMA_EMBED,
     LLM_ARCH_MAINCODER,
     LLM_ARCH_KIMI_LINEAR,
+    LLM_ARCH_KIMI_K3,
     LLM_ARCH_TALKIE,
     LLM_ARCH_MELLUM,
     LLM_ARCH_EAGLE3,
@@ -188,6 +189,9 @@ enum llm_kv {
     LLM_KV_FEATURES_LENGTH,
     LLM_KV_BLOCK_COUNT,
     LLM_KV_LEADING_DENSE_BLOCK_COUNT,
+    LLM_KV_ATTN_RES_BLOCK_SIZE,
+    LLM_KV_ACTIVATION_SITU_BETA,
+    LLM_KV_ACTIVATION_SITU_LINEAR_BETA,
     LLM_KV_FEED_FORWARD_LENGTH,
     LLM_KV_EXPERT_FEED_FORWARD_LENGTH,
     LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH,
@@ -203,6 +207,7 @@ enum llm_kv {
     LLM_KV_EXPERT_GROUP_USED_COUNT,
     LLM_KV_EXPERT_WEIGHTS_SCALE,
     LLM_KV_EXPERT_WEIGHTS_NORM,
+    LLM_KV_EXPERT_LATENT_LENGTH,
     LLM_KV_EXPERT_GATING_FUNC,
     LLM_KV_EXPERT_GROUP_SCALE,
     LLM_KV_EXPERTS_PER_GROUP,
@@ -309,6 +314,7 @@ enum llm_kv {
     LLM_KV_SSM_DT_B_C_RMS,
 
     LLM_KV_KDA_HEAD_DIM,
+    LLM_KV_KDA_GATE_LOWER_BOUND,
 
     LLM_KV_WKV_HEAD_SIZE,
 
@@ -498,6 +504,13 @@ enum llm_tensor {
     LLM_TENSOR_SSM_BETA,            // kimi: beta mixing coefficient and qwen3.5
     LLM_TENSOR_SSM_G_A,             // kimi: output gate projection A
     LLM_TENSOR_SSM_G_B,             // kimi: output gate projection B
+    LLM_TENSOR_SSM_G,               // kimi-k3: full-rank KDA gate
+    LLM_TENSOR_ATTN_RES_SCORE,      // kimi-k3: fused res_norm*res_proj (pre-attn)
+    LLM_TENSOR_FFN_RES_SCORE,       // kimi-k3: fused res_norm*res_proj (pre-ffn)
+    LLM_TENSOR_OUTPUT_RES_SCORE,    // kimi-k3: fused res_norm*res_proj (final)
+    LLM_TENSOR_FFN_ROUTED_DOWN,     // kimi-k3: latent MoE down
+    LLM_TENSOR_FFN_ROUTED_UP,       // kimi-k3: latent MoE up
+    LLM_TENSOR_FFN_ROUTED_NORM,     // kimi-k3: latent MoE norm
     LLM_TENSOR_TIME_MIX_W0,
     LLM_TENSOR_TIME_MIX_W1,
     LLM_TENSOR_TIME_MIX_W2,
diff --git a/src/llama-context.cpp b/src/llama-context.cpp
index 9b399d6096b1..9c7cf59deb86 100644
--- a/src/llama-context.cpp
+++ b/src/llama-context.cpp
@@ -2334,6 +2334,10 @@ void llama_context::output_reorder() {
 //
 
 uint32_t llama_context::graph_max_nodes(uint32_t n_tokens) const {
+    if (model.arch == LLM_ARCH_KIMI_K3) {
+        // the n_tokens*40 budget below is exhausted at ubatch 3840
+        return std::max<uint32_t>(n_tokens * 160, 64u * model.n_tensors());
+    }
     if (model.arch == LLM_ARCH_QWEN3NEXT ||
         model.arch == LLM_ARCH_KIMI_LINEAR ||
         model.arch == LLM_ARCH_QWEN35 ||
diff --git a/src/llama-graph.cpp b/src/llama-graph.cpp
index 980d393db4eb..cf5160b1ea72 100644
--- a/src/llama-graph.cpp
+++ b/src/llama-graph.cpp
@@ -2075,6 +2075,21 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
                 cur = ggml_silu(ctx0, cur);
                 cb(cur, "ffn_moe_silu", il);
             } break;
+        case LLM_FFN_SITU:
+            {
+                // situ(gate, up) = beta*tanh(gate/beta)*sigmoid(gate) * lb*tanh(up/lb)
+                GGML_ASSERT(has_gate);
+                const float beta = hparams.situ_beta;
+                const float lb   = hparams.situ_linear_beta;
+
+                ggml_tensor * act = ggml_scale(ctx0, ggml_tanh(ctx0, ggml_scale(ctx0, cur, 1.0f/beta)), beta);
+                act = ggml_mul(ctx0, act, ggml_sigmoid(ctx0, cur));
+                if (lb > 0.0f) {
+                    up = ggml_scale(ctx0, ggml_tanh(ctx0, ggml_scale(ctx0, up, 1.0f/lb)), lb);
+                }
+                cur = ggml_mul(ctx0, act, up);
+                cb(cur, "ffn_moe_situ", il);
+            } break;
         case LLM_FFN_GELU:
             if (has_gate) {
                 cur = ggml_geglu_split(ctx0, cur, up);
diff --git a/src/llama-graph.h b/src/llama-graph.h
index d3270a1a3469..5e182fbd0468 100644
--- a/src/llama-graph.h
+++ b/src/llama-graph.h
@@ -58,6 +58,7 @@ enum llm_ffn_op_type : int {
     LLM_FFN_GEGLU,
     LLM_FFN_REGLU,
     LLM_FFN_SWIGLU_OAI_MOE,
+    LLM_FFN_SITU,           // kimi-k3 (appended: do not renumber existing values)
 };
 
 enum llm_ffn_gate_type {
diff --git a/src/llama-hparams.h b/src/llama-hparams.h
index c31955700919..0229ae6461df 100644
--- a/src/llama-hparams.h
+++ b/src/llama-hparams.h
@@ -4,10 +4,11 @@
 
 #include <array>
 #include <cassert>
+#include <cmath>
 
 // bump if necessary
 #define LLAMA_MAX_LAYERS  512
-#define LLAMA_MAX_EXPERTS 512 // Qwen3 Next
+#define LLAMA_MAX_EXPERTS 1024 // Kimi K3
 
 enum llama_expert_gating_func_type {
     LLAMA_EXPERT_GATING_FUNC_TYPE_NONE           = 0,
@@ -174,6 +175,13 @@ struct llama_hparams {
     // for Kimi Linear KDA
     uint32_t n_embd_head_kda = 0;
 
+    // kimi-k3
+    uint32_t n_expert_latent      = 0;      // routed_expert_hidden_size (0 = experts run at n_embd)
+    uint32_t attn_res_block_size  = 0;      // 0 = no cross-layer attention residuals
+    float    kda_gate_lower_bound = -INFINITY;
+    float    situ_beta            = 1.0f;
+    float    situ_linear_beta     = 0.0f;   // 0 = no linear-beta transform on the up branch
+
     bool ssm_dt_b_c_rms = false;
 
     float f_clamp_kqv      = 0.0f;
diff --git a/src/llama-model.cpp b/src/llama-model.cpp
index d74a158f4a37..720aba20052d 100644
--- a/src/llama-model.cpp
+++ b/src/llama-model.cpp
@@ -309,6 +309,8 @@ static llama_model * llama_model_mapping(llm_arch arch, const llama_model_params
             return new llama_model_mimo2(params);
         case LLM_ARCH_KIMI_LINEAR:
             return new llama_model_kimi_linear(params);
+        case LLM_ARCH_KIMI_K3:
+            return new llama_model_kimi_k3(params);
         case LLM_ARCH_STEP35:
             return new llama_model_step35(params);
         case LLM_ARCH_INKLING:
@@ -833,6 +835,7 @@ const char * llm_type_name(llm_type type) {
         case LLM_TYPE_397B_A17B:     return "397B.A17B";
         case LLM_TYPE_685B_A37B:     return "685B.A37B";
         case LLM_TYPE_744B_A40B:     return "744B.A40B";
+        case LLM_TYPE_2_8T_A50B:     return "2.8T.A50B";
         case LLM_TYPE_E2B:           return "E2B";
         case LLM_TYPE_E4B:           return "E4B";
         default:                     return "?B";
@@ -2506,6 +2509,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
         case LLM_ARCH_NEMOTRON_H:
         case LLM_ARCH_NEMOTRON_H_MOE:
         case LLM_ARCH_KIMI_LINEAR:
+        case LLM_ARCH_KIMI_K3:
         case LLM_ARCH_INKLING:
             return LLAMA_ROPE_TYPE_NONE;
 
diff --git a/src/llama-model.h b/src/llama-model.h
index 4184111212f3..12f588aa2434 100644
--- a/src/llama-model.h
+++ b/src/llama-model.h
@@ -143,6 +143,7 @@ enum llm_type {
     LLM_TYPE_397B_A17B, // Qwen3.5
     LLM_TYPE_685B_A37B, // DeepSeek V3.2
     LLM_TYPE_744B_A40B, // GLM-5
+    LLM_TYPE_2_8T_A50B, // Kimi-K3
     LLM_TYPE_E2B,
     LLM_TYPE_E4B,
 };
@@ -510,6 +511,14 @@ struct llama_layer {
     struct ggml_tensor * ssm_g_b    = nullptr;
     struct ggml_tensor * ssm_o_norm = nullptr;
 
+    // kimi-k3
+    struct ggml_tensor * ssm_g           = nullptr; // full-rank KDA gate (replaces ssm_g_a/ssm_g_b)
+    struct ggml_tensor * attn_res_score  = nullptr; // fused res_norm*res_proj, pre-attention
+    struct ggml_tensor * ffn_res_score   = nullptr; // fused res_norm*res_proj, pre-FFN
+    struct ggml_tensor * ffn_routed_down = nullptr; // latent MoE: n_embd -> n_expert_latent
+    struct ggml_tensor * ffn_routed_up   = nullptr; // latent MoE: n_expert_latent -> n_embd
+    struct ggml_tensor * ffn_routed_norm = nullptr;
+
     // DSA (deepseek sparse attention)
     struct ggml_tensor * indexer_k_norm   = nullptr;
     struct ggml_tensor * indexer_k_norm_b = nullptr;
@@ -576,6 +585,7 @@ struct llama_model {
     struct ggml_tensor * tok_norm_b = nullptr;
 
     struct ggml_tensor * output_norm     = nullptr;
+    struct ggml_tensor * output_res_score = nullptr; // kimi-k3: final cross-layer residual mix
     struct ggml_tensor * output_norm_b   = nullptr;
     struct ggml_tensor * output          = nullptr;
     struct ggml_tensor * output_b        = nullptr;
diff --git a/src/models/kimi-k3.cpp b/src/models/kimi-k3.cpp
new file mode 100644
index 000000000000..ff7927e3802e
--- /dev/null
+++ b/src/models/kimi-k3.cpp
@@ -0,0 +1,648 @@
+#include "models.h"
+#include "llama-memory-recurrent.h"
+
+//
+// Kimi-K3 text model.
+//
+// Hybrid KDA (linear) + MLA (full) attention, as in Kimi-Linear-48B, plus five
+// things that architecture does not have:
+//
+//   1. cross-layer residual attention  (attn_res_block_size)
+//   2. latent MoE                      (routed experts run at n_expert_latent)
+//   3. situ activation                 (replaces SwiGLU everywhere)
+//   4. MLA output gate                 (sigmoid gate before o_proj)
+//   5. full-rank KDA gate              (single ssm_g instead of ssm_g_a/ssm_g_b)
+//
+
+void llama_model_kimi_k3::load_arch_hparams(llama_model_loader & ml) {
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+    ml.get_key(LLM_KV_ATTENTION_KEY_LENGTH_MLA,    hparams.n_embd_head_k_mla_impl);
+    ml.get_key(LLM_KV_ATTENTION_VALUE_LENGTH_MLA,  hparams.n_embd_head_v_mla_impl);
+    ml.get_key(LLM_KV_ATTENTION_Q_LORA_RANK,       hparams.n_lora_q, false);
+    ml.get_key(LLM_KV_ATTENTION_KV_LORA_RANK,      hparams.n_lora_kv);
+    ml.get_key(LLM_KV_SSM_CONV_KERNEL,             hparams.ssm_d_conv);
+    ml.get_key(LLM_KV_KDA_HEAD_DIM,                hparams.n_embd_head_kda);
+    ml.get_key(LLM_KV_KDA_GATE_LOWER_BOUND,        hparams.kda_gate_lower_bound, false);
+
+    // n_head_kv == 0 marks a KDA (recurrent) layer, as in kimi-linear
+    for (uint32_t i = 0; i < hparams.n_layer(); ++i) {
+        hparams.is_recr_impl[i] = hparams.n_head_kv(i) == 0;
+    }
+
+    ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp);
+    ml.get_key(LLM_KV_EXPERT_SHARED_COUNT,        hparams.n_expert_shared);
+    ml.get_key(LLM_KV_LEADING_DENSE_BLOCK_COUNT,  hparams.n_layer_dense_lead, false);
+    ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE,       hparams.expert_weights_scale, false);
+    ml.get_key(LLM_KV_EXPERT_WEIGHTS_NORM,        hparams.expert_weights_norm, false);
+    ml.get_key(LLM_KV_EXPERT_GATING_FUNC,         hparams.expert_gating_func);
+    // required: a silent default here loads cleanly and produces garbage
+    ml.get_key(LLM_KV_EXPERT_LATENT_LENGTH,         hparams.n_expert_latent);
+    ml.get_key(LLM_KV_ATTN_RES_BLOCK_SIZE,          hparams.attn_res_block_size);
+    ml.get_key(LLM_KV_ACTIVATION_SITU_BETA,         hparams.situ_beta);
+    ml.get_key(LLM_KV_ACTIVATION_SITU_LINEAR_BETA,  hparams.situ_linear_beta);
+
+    GGML_ASSERT(hparams.attn_res_block_size > 0 && "Kimi-K3 requires attn_res.block_size");
+    GGML_ASSERT(hparams.n_expert_latent   > 0 && "Kimi-K3 requires expert_latent_length");
+
+    switch (hparams.n_layer()) {
+        case 93: type = LLM_TYPE_2_8T_A50B; break; // Kimi-K3
+        default: type = LLM_TYPE_UNKNOWN;
+    }
+}
+
+void llama_model_kimi_k3::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+
+    const int64_t n_embd_latent = hparams.n_expert_latent > 0 ? hparams.n_expert_latent : n_embd;
+
+    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, 0);
+
+    if (hparams.attn_res_block_size > 0) {
+        output_res_score = create_tensor(tn(LLM_TENSOR_OUTPUT_RES_SCORE, "weight"), {n_embd}, 0);
+    }
+
+    for (int i = 0; i < n_layer; ++i) {
+        auto & layer = layers[i];
+
+        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+        layer.ffn_norm  = create_tensor(tn(LLM_TENSOR_FFN_NORM,  "weight", i), {n_embd}, 0);
+
+        if (hparams.attn_res_block_size > 0) {
+            layer.attn_res_score = create_tensor(tn(LLM_TENSOR_ATTN_RES_SCORE, "weight", i), {n_embd}, 0);
+            layer.ffn_res_score  = create_tensor(tn(LLM_TENSOR_FFN_RES_SCORE,  "weight", i), {n_embd}, 0);
+        }
+
+        const int64_t head_dim = hparams.n_embd_head_kda;
+        const int64_t d_conv   = hparams.ssm_d_conv;
+        const int64_t d_inner  = head_dim * n_head;
+
+        if (hparams.is_recr(i)) {
+            // conv1d may be stored 4D [d_conv, 1, d_inner, 1] or 3D (quantization drops the trailing 1)
+            auto conv = [&](llm_tensor tid) {
+                ggml_tensor * t = create_tensor(tn(tid, "weight", i), {d_conv, 1, d_inner, 1}, TENSOR_NOT_REQUIRED);
+                return t ? t : create_tensor(tn(tid, "weight", i), {d_conv, 1, d_inner}, 0);
+            };
+            layer.ssm_q_conv = conv(LLM_TENSOR_SSM_CONV1D_Q);
+            layer.ssm_k_conv = conv(LLM_TENSOR_SSM_CONV1D_K);
+            layer.ssm_v_conv = conv(LLM_TENSOR_SSM_CONV1D_V);
+
+            create_tensor_qkv(layer, i, n_embd, d_inner, d_inner, d_inner, 0);
+
+            layer.ssm_f_a  = create_tensor(tn(LLM_TENSOR_SSM_F_A,  "weight", i), {n_embd, head_dim}, 0);
+            layer.ssm_f_b  = create_tensor(tn(LLM_TENSOR_SSM_F_B,  "weight", i), {head_dim, d_inner}, 0);
+            layer.ssm_beta = create_tensor(tn(LLM_TENSOR_SSM_BETA, "weight", i), {n_embd, n_head}, 0);
+
+            // K3's A_log is a plain 1-D [n_head] parameter (kimi-linear's is padded);
+            // accept the padded forms too so both layouts load. -exp() is folded at
+            // conversion time. Only the element count matters - the graph reshapes it.
+            layer.ssm_a = create_tensor(tn(LLM_TENSOR_SSM_A, i), {n_head}, TENSOR_NOT_REQUIRED);
+            if (!layer.ssm_a) {
+                layer.ssm_a = create_tensor(tn(LLM_TENSOR_SSM_A, i), {1, n_head, 1, 1}, TENSOR_NOT_REQUIRED);
+            }
+            if (!layer.ssm_a) {
+                layer.ssm_a = create_tensor(tn(LLM_TENSOR_SSM_A, i), {1, n_head}, 0);
+            }
+
+            layer.ssm_dt_b = create_tensor(tn(LLM_TENSOR_SSM_DT, "bias", i), {d_inner}, 0);
+
+            // K3 uses a single full-rank gate instead of kimi-linear's g_a/g_b pair
+            layer.ssm_g      = create_tensor(tn(LLM_TENSOR_SSM_G,    "weight", i), {n_embd, d_inner}, 0);
+            layer.ssm_o_norm = create_tensor(tn(LLM_TENSOR_SSM_NORM, "weight", i), {head_dim}, 0);
+            layer.wo         = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {d_inner, n_embd}, 0);
+        } else {
+            const int64_t q_lora_rank      = hparams.n_lora_q;
+            const int64_t kv_lora_rank     = hparams.n_lora_kv;
+            const int64_t n_embd_head_k    = hparams.n_embd_head_k_mla();
+            const int64_t n_embd_head_v    = hparams.n_embd_head_v_mla();
+            const int64_t qk_rope_head_dim = hparams.n_rot();
+            const int64_t qk_nope_head_dim = n_embd_head_k - qk_rope_head_dim;
+
+            layer.attn_q_a_norm  = create_tensor(tn(LLM_TENSOR_ATTN_Q_A_NORM,  "weight", i), {q_lora_rank}, TENSOR_NOT_REQUIRED);
+            layer.attn_kv_a_norm = create_tensor(tn(LLM_TENSOR_ATTN_KV_A_NORM, "weight", i), {kv_lora_rank}, 0);
+
+            if (layer.attn_q_a_norm) {
+                layer.wq_a = create_tensor(tn(LLM_TENSOR_ATTN_Q_A, "weight", i), {n_embd, q_lora_rank}, 0);
+                layer.wq_b = create_tensor(tn(LLM_TENSOR_ATTN_Q_B, "weight", i), {q_lora_rank, n_head * n_embd_head_k}, 0);
+            } else {
+                layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_head * n_embd_head_k}, 0);
+            }
+
+            layer.wkv_a_mqa = create_tensor(tn(LLM_TENSOR_ATTN_KV_A_MQA, "weight", i), {n_embd, kv_lora_rank + qk_rope_head_dim}, 0);
+            layer.wkv_b     = create_tensor(tn(LLM_TENSOR_ATTN_KV_B, "weight", i),
+                                            {kv_lora_rank, n_head * (qk_nope_head_dim + n_embd_head_v)},
+                                            TENSOR_NOT_REQUIRED | TENSOR_SKIP_IF_VIRTUAL);
+            if (!layer.wkv_b) {
+                layer.wk_b = create_tensor(tn(LLM_TENSOR_ATTN_K_B, "weight", i), {qk_nope_head_dim, kv_lora_rank, n_head}, 0);
+                layer.wv_b = create_tensor(tn(LLM_TENSOR_ATTN_V_B, "weight", i), {kv_lora_rank, n_embd_head_v, n_head}, 0);
+            }
+
+            // K3: sigmoid output gate applied to the attention output before o_proj
+            layer.wqkv_gate = create_tensor(tn(LLM_TENSOR_ATTN_GATE, "weight", i), {n_embd, n_head * n_embd_head_v}, TENSOR_NOT_REQUIRED);
+
+            layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_head * n_embd_head_v, n_embd}, 0);
+        }
+
+        if (i < (int) hparams.n_layer_dense_lead) {
+            layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0);
+            layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0);
+            layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff}, 0);
+        } else {
+            const int64_t n_ff_exp = hparams.n_ff_exp;
+
+            layer.ffn_gate_inp    = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,    "weight", i), {n_embd, n_expert}, 0);
+            layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias",   i), {n_expert}, 0);
+
+            // routed experts live in the latent space
+            layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd_latent, n_ff_exp, n_expert}, 0);
+            layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp, n_embd_latent, n_expert}, 0);
+            layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {n_embd_latent, n_ff_exp, n_expert}, 0);
+
+            if (hparams.n_expert_latent > 0) {
+                layer.ffn_routed_down = create_tensor(tn(LLM_TENSOR_FFN_ROUTED_DOWN, "weight", i), {n_embd, n_embd_latent}, 0);
+                layer.ffn_routed_up   = create_tensor(tn(LLM_TENSOR_FFN_ROUTED_UP,   "weight", i), {n_embd_latent, n_embd}, 0);
+                layer.ffn_routed_norm = create_tensor(tn(LLM_TENSOR_FFN_ROUTED_NORM, "weight", i), {n_embd_latent}, TENSOR_NOT_REQUIRED);
+            }
+
+            // shared experts stay at n_embd, width = moe_intermediate_size * n_expert_shared
+            const int64_t n_ff_shexp = n_ff_exp * (hparams.n_expert_shared > 0 ? hparams.n_expert_shared : 1);
+            layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, n_ff_shexp}, TENSOR_NOT_REQUIRED);
+            layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {n_ff_shexp, n_embd}, TENSOR_NOT_REQUIRED);
+            layer.ffn_up_shexp   = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {n_embd, n_ff_shexp}, TENSOR_NOT_REQUIRED);
+        }
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_kimi_k3::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
+
+//
+// situ activation:
+//   situ(gate, up) = beta*tanh(gate/beta)*sigmoid(gate) * linear_beta*tanh(up/linear_beta)
+// linear_beta <= 0 disables the transform on the up branch.
+//
+static ggml_tensor * kimi_k3_situ(ggml_context * ctx0, ggml_tensor * gate, ggml_tensor * up,
+                                  float beta, float linear_beta) {
+    ggml_tensor * a = ggml_scale(ctx0, ggml_tanh(ctx0, ggml_scale(ctx0, gate, 1.0f/beta)), beta);
+    a = ggml_mul(ctx0, a, ggml_sigmoid(ctx0, gate));
+
+    if (linear_beta > 0.0f) {
+        up = ggml_scale(ctx0, ggml_tanh(ctx0, ggml_scale(ctx0, up, 1.0f/linear_beta)), linear_beta);
+    }
+    return ggml_mul(ctx0, a, up);
+}
+
+//
+// cross-layer residual attention
+//
+
+void llama_model_kimi_k3::graph::res_push(ggml_tensor * cur, int64_t n_embd, int64_t n_tokens) {
+    ckpts.push_back(ggml_reshape_3d(ctx0, cur, n_embd, 1, n_tokens));
+}
+
+ggml_tensor * llama_model_kimi_k3::graph::res_stack(int64_t n_embd, int64_t n_tokens) {
+    GGML_UNUSED(n_embd);
+    GGML_UNUSED(n_tokens);
+    if (stack_cache_n == (int) ckpts.size()) {
+        return stack_cache;   // set unchanged since the last mix
+    }
+    ggml_tensor * acc = ckpts[0];
+    for (size_t i = 1; i < ckpts.size(); ++i) {
+        acc = ggml_concat(ctx0, acc, ckpts[i], 1);
+    }
+    stack_cache   = acc;
+    stack_cache_n = (int) ckpts.size();
+    return acc;
+}
+
+ggml_tensor * llama_model_kimi_k3::graph::res_mix(ggml_tensor * cur, ggml_tensor * score_w,
+                                                  int64_t n_embd, int64_t n_tokens, int il) {
+    const int n_ckpt = (int) ckpts.size();
+    if (n_ckpt == 0) {
+        return cur; // layer 0: nothing banked yet
+    }
+
+    const float eps = hparams.f_norm_rms_eps;
+
+    ggml_tensor * src = res_stack(n_embd, n_tokens);   // [n_embd, n_ckpt, n_tokens]
+
+    // Scores for the banked checkpoints. One rms_norm covers all of them because
+    // ne0 is n_embd. NOTE: the scores use the *normalized* values, but the weighted
+    // sum below uses the *raw* ones - mirroring _apply_attn_res.
+    ggml_tensor * sc_src = ggml_rms_norm(ctx0, src, eps);
+    sc_src = ggml_mul(ctx0, sc_src, score_w);
+    sc_src = ggml_sum_rows(ctx0, sc_src);                          // [1, n_ckpt, n_tokens]
+    sc_src = ggml_reshape_2d(ctx0, sc_src, n_ckpt, n_tokens);
+
+    // The current residual stream is scored separately and kept out of the stack,
+    // so the stack stays append-only.
+    ggml_tensor * sc_cur = ggml_rms_norm(ctx0, cur, eps);
+    sc_cur = ggml_mul(ctx0, sc_cur, score_w);
+    sc_cur = ggml_sum_rows(ctx0, sc_cur);                          // [1, n_tokens]
+
+    ggml_tensor * scores = ggml_concat(ctx0, sc_src, sc_cur, 0);   // [n_ckpt+1, n_tokens]
+    ggml_tensor * probs  = ggml_soft_max(ctx0, scores);            // over ne0 = n_ckpt+1
+    cb(probs, "res_probs", il);
+
+    // Split the convex combination: hc_pre reduces over ne1 for the stacked part,
+    // a plain broadcast-multiply handles the current stream.
+    ggml_tensor * p_src = ggml_cont(ctx0, ggml_view_2d(ctx0, probs, n_ckpt, n_tokens, probs->nb[1], 0));
+    ggml_tensor * p_cur = ggml_cont(ctx0, ggml_view_2d(ctx0, probs, 1, n_tokens, probs->nb[1],
+                                                       probs->nb[0] * n_ckpt));
+
+    ggml_tensor * out = ggml_dsv4_hc_pre(ctx0, src, p_src);
+    out = ggml_add(ctx0, out, ggml_mul(ctx0, cur, p_cur));
+
+    return out;
+}
+
+llama_model_kimi_k3::graph::graph(const llama_model & model, const llm_graph_params & params) :
+    llm_build_delta_net_base(params), model(model) {
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+    cb(inpL, "inp_embd", -1);
+
+    // K3 MLA is nope-only, so there is no position input
+
+    auto * inp_kv      = !hparams.is_mla() ? build_inp_mem_hybrid()   : nullptr;
+    auto * inp_k       =  hparams.is_mla() ? build_inp_mem_hybrid_k() : nullptr;
+    auto * inp_rs      =  hparams.is_mla() ? inp_k->get_recr() : inp_kv->get_recr();
+    auto * inp_attn_kv = !hparams.is_mla() ? inp_kv->get_attn() : nullptr;
+    auto * inp_attn_k  =  hparams.is_mla() ? inp_k->get_attn()  : nullptr;
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    const int64_t n_head_kda = hparams.n_head();
+    const int64_t head_dim   = hparams.n_embd_head_kda;
+    const int64_t d_conv     = hparams.ssm_d_conv;
+    const int64_t d_inner    = n_head_kda * head_dim;
+    const int64_t n_seqs     = ubatch.n_seqs;
+    const int64_t n_seq_tokens = ubatch.n_seq_tokens;
+
+    GGML_ASSERT(n_seqs != 0);
+    GGML_ASSERT(ubatch.equal_seqs());
+    GGML_ASSERT(ubatch.n_tokens == n_seq_tokens * n_seqs);
+
+    const int64_t n_embd_head_k_mla   = hparams.n_embd_head_k_mla();
+    const int64_t n_embd_head_v_mla   = hparams.n_embd_head_v_mla();
+    const int64_t kv_lora_rank        = hparams.n_lora_kv;
+    const int64_t n_embd_head_qk_rope = hparams.n_rot();
+    const int64_t n_embd_head_qk_nope = n_embd_head_k_mla - n_embd_head_qk_rope;
+    const float   kq_scale_mla        = 1.0f / sqrtf((float) n_embd_head_k_mla);
+
+    const uint32_t res_bs        = hparams.attn_res_block_size;
+    const bool     use_attn_res  = res_bs > 0;
+    const int64_t  n_embd_latent = hparams.n_expert_latent > 0 ? hparams.n_expert_latent : n_embd;
+
+    for (int il = 0; il < n_layer; ++il) {
+        const auto & layer = model.layers[il];
+
+        // `prefix_sum` is the residual stream. On checkpoint layers it is banked
+        // into res_stack and restarts from the attention output alone.
+        ggml_tensor * prefix_sum = inpL;
+
+        cur = use_attn_res ? res_mix(prefix_sum, layer.attn_res_score, n_embd, n_tokens, il)
+                           : prefix_sum;
+
+        bool banked = false;
+        if (use_attn_res && (uint32_t) il % res_bs == 0) {
+            res_push(prefix_sum, n_embd, n_tokens);  // banks the RAW layer input, not `cur`
+            banked = true;
+        }
+
+        cur = build_norm(cur, layer.attn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+        ggml_build_forward_expand(gf, cur);
+
+        if (hparams.is_recr(il)) {
+            cur = build_kda_layer(cur, layer, inp_rs, d_conv, head_dim, n_head_kda,
+                                  d_inner, n_seq_tokens, n_seqs, il);
+        } else {
+            cur = build_mla_layer(cur, layer, inp_attn_k, inp_attn_kv,
+                                  n_embd_head_k_mla, n_embd_head_v_mla, kv_lora_rank,
+                                  n_embd_head_qk_rope, n_embd_head_qk_nope, kq_scale_mla, il);
+        }
+
+        prefix_sum = banked ? cur : ggml_add(ctx0, prefix_sum, cur);
+        cb(prefix_sum, "prefix_sum_attn", il);
+
+        cur = use_attn_res ? res_mix(prefix_sum, layer.ffn_res_score, n_embd, n_tokens, il)
+                           : prefix_sum;
+
+        cur = build_norm(cur, layer.ffn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        if ((uint32_t) il < hparams.n_layer_dense_lead) {
+            ggml_tensor * g = ggml_mul_mat(ctx0, layer.ffn_gate, cur);
+            ggml_tensor * u = ggml_mul_mat(ctx0, layer.ffn_up,   cur);
+            cur = kimi_k3_situ(ctx0, g, u, hparams.situ_beta, hparams.situ_linear_beta);
+            cur = ggml_mul_mat(ctx0, layer.ffn_down, cur);
+            cb(cur, "ffn_out", il);
+        } else {
+            cur = build_latent_moe(cur, layer, n_embd_latent, il);
+        }
+
+        prefix_sum = ggml_add(ctx0, prefix_sum, cur);
+        prefix_sum = build_cvec(prefix_sum, il);
+        cb(prefix_sum, "l_out", il);
+
+        inpL = prefix_sum;
+    }
+
+    cur = inpL;
+
+    // final mix, then narrow to the output tokens
+    if (use_attn_res) {
+        cur = res_mix(cur, model.output_res_score, n_embd, n_tokens, -1);
+    }
+    if (inp_out_ids) {
+        cur = ggml_get_rows(ctx0, cur, inp_out_ids);
+    }
+
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    cur = ggml_mul_mat(ctx0, model.output, cur);
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
+
+//
+// KDA layer
+//
+
+// Causal conv1d over one of Q/K/V. `qkv` selects which third of the conv state to use.
+// Polled rather than blocking is not a concern here; this mirrors kimi-linear's helper.
+static ggml_tensor * kimi_k3_conv1d(ggml_cgraph * gf, ggml_context * ctx0,
+                                    ggml_tensor * conv_states_all, ggml_tensor * conv_state_all,
+                                    int64_t qkv, ggml_tensor * x, ggml_tensor * proj_w, ggml_tensor * conv_w,
+                                    int64_t d_conv, int64_t head_dim, int64_t n_head,
+                                    int64_t n_seq_tokens, int64_t n_seqs, int64_t n_tokens, int64_t kv_head) {
+    const int64_t d_inner         = head_dim * n_head;
+    const int64_t conv_state_size = (d_conv - 1) * d_inner;
+    const int64_t n_embd_r_total  = 3 * conv_state_size;
+
+    ggml_tensor * conv_state_x = ggml_view_3d(ctx0, conv_state_all, d_conv - 1, d_inner, n_seqs,
+        (d_conv - 1)   * ggml_element_size(conv_state_all),
+        n_embd_r_total * ggml_element_size(conv_state_all),
+        qkv * conv_state_size * ggml_element_size(conv_state_all));
+
+    ggml_tensor * x_proj = ggml_mul_mat(ctx0, proj_w, x);
+    ggml_tensor * x_3d   = ggml_reshape_3d(ctx0, x_proj, d_inner, n_seq_tokens, n_seqs);
+    ggml_tensor * conv_x = ggml_concat(ctx0, conv_state_x, ggml_transpose(ctx0, x_3d), 0);
+
+    ggml_tensor * last_conv_x = ggml_view_3d(ctx0, conv_x, d_conv - 1, d_inner, n_seqs,
+        conv_x->nb[1], conv_x->nb[2], n_seq_tokens * conv_x->nb[0]);
+    ggml_build_forward_expand(gf,
+        ggml_cpy(ctx0, last_conv_x,
+            ggml_view_3d(ctx0, conv_states_all, d_conv - 1, d_inner, n_seqs,
+                (d_conv - 1)   * ggml_element_size(conv_states_all),
+                n_embd_r_total * ggml_element_size(conv_states_all),
+                (kv_head * n_embd_r_total + qkv * conv_state_size) * ggml_element_size(conv_states_all))));
+
+    ggml_tensor * conv_weight = ggml_reshape_2d(ctx0, conv_w, d_conv, d_inner);
+    ggml_tensor * Xcur = ggml_ssm_conv(ctx0, conv_x, conv_weight);
+    Xcur = ggml_reshape_2d(ctx0, Xcur, d_inner, n_tokens);
+    Xcur = ggml_silu(ctx0, Xcur);
+
+    return ggml_reshape_4d(ctx0, Xcur, head_dim, n_head, n_seq_tokens, n_seqs);
+}
+
+ggml_tensor * llama_model_kimi_k3::graph::build_kda_layer(
+        ggml_tensor * cur, const llama_layer & layer, llm_graph_input_rs * inp_rs,
+        int64_t d_conv, int64_t head_dim, int64_t n_head_kda,
+        int64_t d_inner, int64_t n_seq_tokens, int64_t n_seqs, int il) {
+
+    const auto * mctx_cur = inp_rs->mctx;
+    const auto   kv_head  = mctx_cur->get_head();
+
+    ggml_tensor * conv_states_all = mctx_cur->get_r_l(il);
+    ggml_tensor * conv_state_all  = build_rs(inp_rs, conv_states_all, hparams.n_embd_r(), n_seqs);
+
+    ggml_tensor * Qcur = kimi_k3_conv1d(gf, ctx0, conv_states_all, conv_state_all, 0, cur, layer.wq, layer.ssm_q_conv, d_conv, head_dim, n_head_kda, n_seq_tokens, n_seqs, n_tokens, kv_head);
+    ggml_tensor * Kcur = kimi_k3_conv1d(gf, ctx0, conv_states_all, conv_state_all, 1, cur, layer.wk, layer.ssm_k_conv, d_conv, head_dim, n_head_kda, n_seq_tokens, n_seqs, n_tokens, kv_head);
+    ggml_tensor * Vcur = kimi_k3_conv1d(gf, ctx0, conv_states_all, conv_state_all, 2, cur, layer.wv, layer.ssm_v_conv, d_conv, head_dim, n_head_kda, n_seq_tokens, n_seqs, n_tokens, kv_head);
+    cb(Qcur, "kda_q_conv", il);
+    cb(Kcur, "kda_k_conv", il);
+    cb(Vcur, "kda_v_conv", il);
+
+    // The decay gate has two forms, selected by linear_attn_config.gate_lower_bound
+    // (fla/ops/kda/gate.py). `lower_bound` is NOT a clamp - when set it swaps the
+    // activation entirely:
+    //
+    //   unset (kimi-linear):  g = -exp(A_log) * softplus(f_b(f_a(x)) + dt_bias)
+    //   set   (K3, -5.0):     g = lower_bound * sigmoid(exp(A_log) * (f_b(f_a(x)) + dt_bias))
+    //
+    // ssm_a holds -exp(A_log) (folded at conversion time), so exp(A_log) == -ssm_a.
+    ggml_tensor * f_a = ggml_mul_mat(ctx0, layer.ssm_f_a, cur);
+    ggml_tensor * g1  = ggml_mul_mat(ctx0, layer.ssm_f_b, f_a);
+    g1 = ggml_add(ctx0, g1, layer.ssm_dt_b);
+
+    ggml_tensor * A = ggml_reshape_3d(ctx0, layer.ssm_a, 1, n_head_kda, 1);
+
+    if (hparams.kda_gate_lower_bound > -INFINITY) {
+        g1 = ggml_reshape_3d(ctx0, g1, head_dim, n_head_kda, n_tokens);
+        g1 = ggml_mul(ctx0, g1, A);                                    // -exp(A_log) * (...)
+        g1 = ggml_sigmoid(ctx0, ggml_scale(ctx0, g1, -1.0f));
+        g1 = ggml_scale(ctx0, g1, hparams.kda_gate_lower_bound);
+    } else {
+        g1 = ggml_softplus(ctx0, g1);
+        g1 = ggml_reshape_3d(ctx0, g1, head_dim, n_head_kda, n_tokens);
+        g1 = ggml_mul(ctx0, g1, A);
+    }
+    cb(g1, "kda_g1", il);
+
+    g1 = ggml_reshape_4d(ctx0, g1, head_dim, n_head_kda, n_seq_tokens, n_seqs);
+
+    ggml_tensor * beta = ggml_mul_mat(ctx0, layer.ssm_beta, cur);
+    beta = ggml_reshape_4d(ctx0, beta, 1, n_head_kda, n_seq_tokens, n_seqs);
+    beta = ggml_sigmoid(ctx0, beta);
+    cb(beta, "kda_beta", il);
+
+    ggml_tensor * cur_3d = ggml_reshape_3d(ctx0, cur, cur->ne[0], n_seq_tokens, n_seqs);
+
+    ggml_tensor * ssm_states_all = mctx_cur->get_s_l(il);
+    ggml_tensor * state = build_rs(inp_rs, ssm_states_all, hparams.n_embd_s(), n_seqs);
+    state = ggml_reshape_4d(ctx0, state, head_dim, head_dim, n_head_kda, n_seqs);
+
+    const float eps = hparams.f_norm_rms_eps;
+    Qcur = ggml_l2_norm(ctx0, Qcur, eps);
+    Kcur = ggml_l2_norm(ctx0, Kcur, eps);
+
+    auto attn_out = build_delta_net(Qcur, Kcur, Vcur, g1, beta, state, il);
+
+    ggml_tensor * output    = ggml_cont(ctx0, attn_out.first);
+    cb(output, "kda_scan_out", il);
+    ggml_tensor * new_state = attn_out.second;
+
+    ggml_build_forward_expand(gf,
+        ggml_cpy(ctx0, new_state,
+            ggml_view_1d(ctx0, ssm_states_all, hparams.n_embd_s() * n_seqs,
+                         kv_head * hparams.n_embd_s() * ggml_element_size(ssm_states_all))));
+
+    // K3: single full-rank gate (kimi-linear factors this as g_b(g_a(x)))
+    ggml_tensor * cur_2d = ggml_reshape_2d(ctx0, cur_3d, cur_3d->ne[0], n_seq_tokens * n_seqs);
+    ggml_tensor * g2     = ggml_mul_mat(ctx0, layer.ssm_g, cur_2d);
+    g2 = ggml_reshape_3d(ctx0, g2, head_dim, n_head_kda, n_seq_tokens * n_seqs);
+
+    ggml_tensor * o      = ggml_reshape_3d(ctx0, output, head_dim, n_head_kda, n_seq_tokens * n_seqs);
+    ggml_tensor * normed = build_norm(o, layer.ssm_o_norm, nullptr, LLM_NORM_RMS, il);
+    cb(g2, "kda_g2", il);
+    cb(normed, "kda_normed", il);
+    ggml_tensor * gated  = ggml_mul(ctx0, normed, ggml_sigmoid(ctx0, g2));
+
+    gated = ggml_cont_2d(ctx0, gated, d_inner, n_tokens);
+    cur   = ggml_mul_mat(ctx0, layer.wo, gated);
+    cb(cur, "kda_out", il);
+
+    return cur;
+}
+
+//
+// MLA layer (nope-only, with K3's sigmoid output gate)
+//
+
+ggml_tensor * llama_model_kimi_k3::graph::build_mla_layer(
+        ggml_tensor * cur, const llama_layer & layer,
+        llm_graph_input_attn_k * inp_attn_k, llm_graph_input_attn_kv * inp_attn_kv,
+        int64_t n_embd_head_k_mla, int64_t n_embd_head_v_mla, int64_t kv_lora_rank,
+        int64_t n_embd_head_qk_rope, int64_t n_embd_head_qk_nope, float kq_scale, int il) {
+
+    ggml_tensor * inp_gate = cur; // the output gate reads the *normed* layer input
+
+    ggml_tensor * Qcur;
+    if (layer.wq_a) {
+        Qcur = ggml_mul_mat(ctx0, layer.wq_a, cur);
+        Qcur = build_norm(Qcur, layer.attn_q_a_norm, nullptr, LLM_NORM_RMS, il);
+        Qcur = ggml_mul_mat(ctx0, layer.wq_b, Qcur);
+    } else {
+        Qcur = ggml_mul_mat(ctx0, layer.wq, cur);
+    }
+
+    ggml_tensor * kv_cmpr_pe = ggml_mul_mat(ctx0, layer.wkv_a_mqa, cur);
+
+    ggml_tensor * kv_cmpr = ggml_view_2d(ctx0, kv_cmpr_pe, kv_lora_rank, n_tokens,
+        ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope), 0);
+    ggml_tensor * k_pe = ggml_view_3d(ctx0, kv_cmpr_pe, n_embd_head_qk_rope, 1, n_tokens,
+        ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope),
+        ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope),
+        ggml_row_size(kv_cmpr_pe->type, kv_lora_rank));
+
+    // no RoPE: mla_use_nope is asserted at conversion time
+    kv_cmpr = build_norm(kv_cmpr, layer.attn_kv_a_norm, nullptr, LLM_NORM_RMS, il);
+
+    ggml_tensor * out;
+    if (layer.wk_b && layer.wv_b) {
+        ggml_tensor * q_nope = ggml_view_3d(ctx0, Qcur, n_embd_head_qk_nope, n_head, n_tokens,
+            ggml_row_size(Qcur->type, n_embd_head_k_mla),
+            ggml_row_size(Qcur->type, n_embd_head_k_mla) * n_head, 0);
+        ggml_tensor * q_pe = ggml_view_3d(ctx0, Qcur, n_embd_head_qk_rope, n_head, n_tokens,
+            ggml_row_size(Qcur->type, n_embd_head_k_mla),
+            ggml_row_size(Qcur->type, n_embd_head_k_mla) * n_head,
+            ggml_row_size(Qcur->type, n_embd_head_qk_nope));
+
+        q_nope = ggml_permute(ctx0, q_nope, 0, 2, 1, 3);
+        ggml_tensor * q_nope_absorbed = ggml_mul_mat(ctx0, layer.wk_b, q_nope);
+        q_nope_absorbed = ggml_permute(ctx0, q_nope_absorbed, 0, 2, 1, 3);
+
+        ggml_tensor * Q = ggml_concat(ctx0, q_nope_absorbed, q_pe, 0);
+        ggml_tensor * kv_cmpr_3d = ggml_reshape_3d(ctx0, kv_cmpr, kv_lora_rank, 1, n_tokens);
+        ggml_tensor * K = ggml_concat(ctx0, kv_cmpr_3d, k_pe, 0);
+        ggml_tensor * V = kv_cmpr_3d;
+
+        // wo == NULL: the output projection is applied after the gate below
+        out = build_attn(inp_attn_k, nullptr, NULL, nullptr, Q, K, V, nullptr, nullptr, layer.wv_b, kq_scale, il);
+    } else {
+        ggml_tensor * Q = ggml_reshape_3d(ctx0, Qcur, n_embd_head_k_mla, n_head, n_tokens);
+        ggml_tensor * kv = ggml_mul_mat(ctx0, layer.wkv_b, kv_cmpr);
+        const int64_t kv_per_head = n_embd_head_qk_nope + n_embd_head_v_mla;
+
+        ggml_tensor * k_nope = ggml_view_3d(ctx0, kv, n_embd_head_qk_nope, n_head, n_tokens,
+            ggml_row_size(kv->type, kv_per_head), ggml_row_size(kv->type, kv_per_head * n_head), 0);
+        ggml_tensor * V = ggml_cont(ctx0, ggml_view_3d(ctx0, kv, n_embd_head_v_mla, n_head, n_tokens,
+            ggml_row_size(kv->type, kv_per_head), ggml_row_size(kv->type, kv_per_head * n_head),
+            ggml_row_size(kv->type, n_embd_head_qk_nope)));
+
+        ggml_tensor * k_pe_t = ggml_new_tensor_3d(ctx0, k_pe->type, n_embd_head_qk_rope, n_head, n_tokens);
+        ggml_tensor * K = ggml_concat(ctx0, ggml_repeat(ctx0, k_pe, k_pe_t), k_nope, 0);
+
+        out = build_attn(inp_attn_kv, nullptr, NULL, nullptr, Q, K, V, nullptr, nullptr, nullptr, kq_scale, il);
+    }
+
+    // K3: attn_output *= sigmoid(g_proj(x)), then o_proj
+    if (layer.wqkv_gate) {
+        ggml_tensor * g = ggml_sigmoid(ctx0, ggml_mul_mat(ctx0, layer.wqkv_gate, inp_gate));
+        out = ggml_mul(ctx0, out, g);
+        cb(out, "mla_gated", il);
+    }
+
+    out = ggml_mul_mat(ctx0, layer.wo, out);
+    cb(out, "mla_out", il);
+
+    return out;
+}
+
+//
+// latent MoE: down-project, run the routed experts in the latent space, norm, up-project;
+// shared experts stay at n_embd and read the un-projected input.
+//
+
+ggml_tensor * llama_model_kimi_k3::graph::build_latent_moe(
+        ggml_tensor * cur, const llama_layer & layer, int64_t n_embd_latent, int il) {
+
+    ggml_tensor * identity = cur;
+
+    ggml_tensor * routed_in = layer.ffn_routed_down
+        ? ggml_mul_mat(ctx0, layer.ffn_routed_down, cur)
+        : cur;
+
+    // The router scores the FULL-WIDTH input while the experts consume the latent
+    // one, so the logits are computed here and handed to build_moe_ffn via
+    // `probs_in` (which substitutes for the gate_inp matmul).
+    ggml_tensor * logits = ggml_mul_mat(ctx0, layer.ffn_gate_inp, identity);
+    cb(logits, "ffn_moe_logits", il);
+
+    ggml_tensor * moe_out = build_moe_ffn(routed_in,
+        nullptr, // gate_inp unused: logits supplied below
+        layer.ffn_up_exps,
+        layer.ffn_gate_exps,
+        layer.ffn_down_exps,
+        layer.ffn_exp_probs_b,
+        hparams.n_expert,
+        hparams.n_expert_used,
+        LLM_FFN_SITU, hparams.expert_weights_norm,
+        hparams.expert_weights_scale,
+        (llama_expert_gating_func_type) hparams.expert_gating_func,
+        il,
+        logits);
+    cb(moe_out, "ffn_moe_out", il);
+
+    if (layer.ffn_routed_norm) {
+        moe_out = build_norm(moe_out, layer.ffn_routed_norm, NULL, LLM_NORM_RMS, il);
+    }
+    if (layer.ffn_routed_up) {
+        moe_out = ggml_mul_mat(ctx0, layer.ffn_routed_up, moe_out);
+    }
+    GGML_UNUSED(n_embd_latent);
+
+    if (layer.ffn_gate_shexp) {
+        ggml_tensor * g = ggml_mul_mat(ctx0, layer.ffn_gate_shexp, identity);
+        ggml_tensor * u = ggml_mul_mat(ctx0, layer.ffn_up_shexp,   identity);
+        ggml_tensor * sh = kimi_k3_situ(ctx0, g, u, hparams.situ_beta, hparams.situ_linear_beta);
+        sh = ggml_mul_mat(ctx0, layer.ffn_down_shexp, sh);
+        cb(sh, "ffn_shexp", il);
+        moe_out = ggml_add(ctx0, moe_out, sh);
+    }
+
+    cb(moe_out, "ffn_out", il);
+    return moe_out;
+}
diff --git a/src/models/models.h b/src/models/models.h
index a11cd7424789..f68d7ea234b7 100644
--- a/src/models/models.h
+++ b/src/models/models.h
@@ -2144,6 +2144,57 @@ struct llama_model_mimo2 : public llama_model_base {
 };
 
 
+struct llama_model_kimi_k3 : public llama_model_base {
+    llama_model_kimi_k3(const struct llama_model_params & params) : llama_model_base(params) {}
+    void load_arch_hparams(llama_model_loader & ml) override;
+    void load_arch_tensors(llama_model_loader & ml) override;
+
+    struct graph : public llm_build_delta_net_base {
+        graph(const llama_model & model, const llm_graph_params & params);
+
+        const llama_model & model;
+
+        // Cross-layer residual attention (K3's `_apply_attn_res`).
+        //
+        // `src` holds the block-residual checkpoints, one [n_embd, n_token] slab
+        // per checkpoint, packed as [n_embd, n_ckpt, n_token] so that:
+        //   - ggml_rms_norm reduces over ne0 = n_embd (scores all slabs at once)
+        //   - ggml_dsv4_hc_pre reduces over ne1 = n_ckpt (the weighted sum)
+        // Both requirements are satisfied by the same layout, which is why this
+        // needs no transpose of the (large) stack.
+        // Each checkpoint is kept as its own [n_embd, 1, n_token] tensor and the
+        // [n_embd, n_ckpt, n_token] stack is materialised with ggml_concat only when
+        // the set changes (8 times per forward pass for the real model). A single
+        // preallocated buffer would be cheaper, but a bare ggml_new_tensor leaf has
+        // no backing buffer - ggml-alloc only allocates tensors that are op outputs.
+        std::vector<ggml_tensor *> ckpts;
+        ggml_tensor * stack_cache   = nullptr;
+        int           stack_cache_n = -1;
+
+        void          res_push(ggml_tensor * cur, int64_t n_embd, int64_t n_tokens);
+        ggml_tensor * res_stack(int64_t n_embd, int64_t n_tokens);
+        ggml_tensor * res_mix(ggml_tensor * cur, ggml_tensor * score_w,
+                              int64_t n_embd, int64_t n_tokens, int il);
+
+        ggml_tensor * build_kda_layer(ggml_tensor * cur, const llama_layer & layer,
+                                      llm_graph_input_rs * inp_rs,
+                                      int64_t d_conv, int64_t head_dim, int64_t n_head_kda,
+                                      int64_t d_inner, int64_t n_seq_tokens, int64_t n_seqs, int il);
+
+        ggml_tensor * build_mla_layer(ggml_tensor * cur, const llama_layer & layer,
+                                      llm_graph_input_attn_k  * inp_attn_k,
+                                      llm_graph_input_attn_kv * inp_attn_kv,
+                                      int64_t n_embd_head_k_mla, int64_t n_embd_head_v_mla,
+                                      int64_t kv_lora_rank, int64_t n_embd_head_qk_rope,
+                                      int64_t n_embd_head_qk_nope, float kq_scale, int il);
+
+        ggml_tensor * build_latent_moe(ggml_tensor * cur, const llama_layer & layer,
+                                       int64_t n_embd_latent, int il);
+    };
+
+    std::unique_ptr<llm_graph_context> build_arch_graph(const llm_graph_params & params) const override;
+};
+
 struct llama_model_kimi_linear : public llama_model_base {
     llama_model_kimi_linear(const struct llama_model_params & params) : llama_model_base(params) {}
     void load_arch_hparams(llama_model_loader & ml) override;
diff --git a/tests/test-chat.cpp b/tests/test-chat.cpp
index d0f985049c89..08d69a1b1a12 100644
--- a/tests/test-chat.cpp
+++ b/tests/test-chat.cpp
@@ -4482,6 +4482,111 @@ static void test_template_output_peg_parsers(bool detailed_debug) {
         }
     }
 
+    // Kimi-K3 tests - custom parser
+    // Unique feature: XTML-ish tags built from <|open|>/<|close|>/<|sep|>, and a
+    // generation prompt that leaves the think section already open.
+    {
+        auto tst = peg_tester("models/templates/Kimi-K3.jinja", detailed_debug);
+
+        // Content only. The response section is explicit even with no reasoning.
+        tst.test("<|open|>response<|sep|>Hello, world!\nWhat's up?<|close|>response<|sep|>"
+                 "<|close|>message<|sep|>")
+            .expect(message_assist)
+            .run();
+
+        // Reasoning with NO opening tag - the generation prompt already opened
+        // it. This is the case that silently loses reasoning if unhandled.
+        tst.test("I'm thinking about this<|close|>think<|sep|>"
+                 "<|open|>response<|sep|>Hello, world!\nWhat's up?<|close|>response<|sep|>"
+                 "<|close|>message<|sep|>")
+            .reasoning_format(COMMON_REASONING_FORMAT_AUTO)
+            .expect(simple_assist_msg("Hello, world!\nWhat's up?", "I'm thinking about this"))
+            .run();
+
+        // Prose that mentions the tag names must survive intact.
+        tst.test("<|open|>response<|sep|>Use the response tag, then message the handler."
+                 "<|close|>response<|sep|><|close|>message<|sep|>")
+            .expect(simple_assist_msg("Use the response tag, then message the handler."))
+            .run();
+
+        // Truncated mid-reasoning (hit the token budget): keep the reasoning.
+        tst.test("I was still thinking when the budget ran out")
+            .reasoning_format(COMMON_REASONING_FORMAT_AUTO)
+            .expect_reasoning("I was still thinking when the budget ran out")
+            .run();
+
+        // Single tool call, one argument.
+        tst.test("<|open|>response<|sep|><|close|>response<|sep|>"
+                 "<|open|>tools<|sep|>"
+                 "<|open|>call tool=\"special_function\" index=\"1\"<|sep|>"
+                 "<|open|>argument key=\"arg1\" type=\"number\"<|sep|>1<|close|>argument<|sep|>"
+                 "<|close|>call<|sep|><|close|>tools<|sep|><|close|>message<|sep|>")
+            .tools({ special_function_tool })
+            .expect_tool_calls({
+                { "special_function", R"({"arg1":1})", "" },
+            })
+            .run();
+
+        // Tool call preceded by reasoning (no opening think tag) and content.
+        tst.test("I should call it<|close|>think<|sep|>"
+                 "<|open|>response<|sep|>On it.<|close|>response<|sep|>"
+                 "<|open|>tools<|sep|>"
+                 "<|open|>call tool=\"special_function\" index=\"1\"<|sep|>"
+                 "<|open|>argument key=\"arg1\" type=\"number\"<|sep|>1<|close|>argument<|sep|>"
+                 "<|close|>call<|sep|><|close|>tools<|sep|><|close|>message<|sep|>")
+            .reasoning_format(COMMON_REASONING_FORMAT_AUTO)
+            .tools({ special_function_tool })
+            .expect(simple_assist_msg("On it.", "I should call it", "special_function",
+                                      R"({"arg1":1})", ""))
+            .run();
+
+        // Multiple typed arguments: the type lives in an attribute, and the
+        // value must come back as a JSON number, not the string "2".
+        tst.test("<|open|>response<|sep|><|close|>response<|sep|>"
+                 "<|open|>tools<|sep|>"
+                 "<|open|>call tool=\"special_function_with_opt\" index=\"1\"<|sep|>"
+                 "<|open|>argument key=\"arg1\" type=\"number\"<|sep|>1<|close|>argument<|sep|>"
+                 "<|open|>argument key=\"arg2\" type=\"number\"<|sep|>2<|close|>argument<|sep|>"
+                 "<|close|>call<|sep|><|close|>tools<|sep|><|close|>message<|sep|>")
+            .tools({ special_function_tool_with_optional_param })
+            .expect_tool_calls({
+                { "special_function_with_opt", R"({"arg1":1,"arg2":2})", "" },
+            })
+            .run();
+
+        // Parallel tool calls in one <|open|>tools<|sep|> section.
+        tst.test("<|open|>response<|sep|><|close|>response<|sep|>"
+                 "<|open|>tools<|sep|>"
+                 "<|open|>call tool=\"special_function\" index=\"1\"<|sep|>"
+                 "<|open|>argument key=\"arg1\" type=\"number\"<|sep|>1<|close|>argument<|sep|>"
+                 "<|close|>call<|sep|>"
+                 "<|open|>call tool=\"special_function_with_opt\" index=\"2\"<|sep|>"
+                 "<|open|>argument key=\"arg1\" type=\"number\"<|sep|>1<|close|>argument<|sep|>"
+                 "<|open|>argument key=\"arg2\" type=\"number\"<|sep|>2<|close|>argument<|sep|>"
+                 "<|close|>call<|sep|><|close|>tools<|sep|><|close|>message<|sep|>")
+            .parallel_tool_calls(true)
+            .tools({ special_function_tool, special_function_tool_with_optional_param })
+            .expect_tool_calls({
+                { "special_function", R"({"arg1":1})", "" },
+                { "special_function_with_opt", R"({"arg1":1,"arg2":2})", "" },
+            })
+            .run();
+
+        // String-typed argument keeps its literal text (no JSON coercion).
+        tst.test("<|open|>response<|sep|><|close|>response<|sep|>"
+                 "<|open|>tools<|sep|>"
+                 "<|open|>call tool=\"python\" index=\"1\"<|sep|>"
+                 "<|open|>argument key=\"code\" type=\"string\"<|sep|>print('hey')"
+                 "<|close|>argument<|sep|>"
+                 "<|close|>call<|sep|><|close|>tools<|sep|><|close|>message<|sep|>")
+            .tools({ python_tool })
+            .expect_tool_calls({
+                // custom delimiter: the payload itself contains )"
+                { "python", R"JSON({"code":"print('hey')"})JSON", "" },
+            })
+            .run();
+    }
+
     // Kimi-K2-Thinking tests - custom parser
     // Unique feature: tool call ID embeds function name as functions.<name>:<counter>
     {
diff --git a/tests/test-llama-archs.cpp b/tests/test-llama-archs.cpp
index 792617731a04..52a73656dbbb 100644
--- a/tests/test-llama-archs.cpp
+++ b/tests/test-llama-archs.cpp
@@ -105,6 +105,7 @@ static gguf_context_ptr get_gguf_ctx(const llm_arch arch, const bool moe) {
             || arch == LLM_ARCH_DEEPSEEK32
             || arch == LLM_ARCH_GLM_DSA
             || arch == LLM_ARCH_KIMI_LINEAR
+            || arch == LLM_ARCH_KIMI_K3
             || arch == LLM_ARCH_MISTRAL4) {
         n_embd = 128;
         n_head = 1;
@@ -143,7 +144,7 @@ static gguf_context_ptr get_gguf_ctx(const llm_arch arch, const bool moe) {
     ms.add_kv(LLM_KV_FULL_ATTENTION_INTERVAL, uint32_t(2));
 
     if (arch == LLM_ARCH_PLAMO2 || arch == LLM_ARCH_JAMBA || arch == LLM_ARCH_NEMOTRON_H || arch == LLM_ARCH_NEMOTRON_H_MOE ||
-            arch == LLM_ARCH_GRANITE_HYBRID || arch == LLM_ARCH_LFM2 || arch == LLM_ARCH_LFM2MOE || arch == LLM_ARCH_KIMI_LINEAR) {
+            arch == LLM_ARCH_GRANITE_HYBRID || arch == LLM_ARCH_LFM2 || arch == LLM_ARCH_LFM2MOE || arch == LLM_ARCH_KIMI_LINEAR || arch == LLM_ARCH_KIMI_K3) {
         GGML_ASSERT(n_layer >= 2);
         std::vector<uint32_t> n_head_per_layer;
         n_head_per_layer.reserve(n_layer);
@@ -162,6 +163,7 @@ static gguf_context_ptr get_gguf_ctx(const llm_arch arch, const bool moe) {
             || arch == LLM_ARCH_DEEPSEEK32
             || arch == LLM_ARCH_GLM_DSA
             || arch == LLM_ARCH_KIMI_LINEAR
+            || arch == LLM_ARCH_KIMI_K3
             || arch == LLM_ARCH_MISTRAL4) {
         ms.add_kv(LLM_KV_ATTENTION_KEY_LENGTH,       uint32_t(576));
         ms.add_kv(LLM_KV_ATTENTION_VALUE_LENGTH,     uint32_t(512));
@@ -367,6 +369,7 @@ static bool moe_mandatory(const llm_arch arch) {
         case LLM_ARCH_PADDLEOCR:
         case LLM_ARCH_MIMO2:
         case LLM_ARCH_KIMI_LINEAR:
+        case LLM_ARCH_KIMI_K3:
         case LLM_ARCH_STEP35:
         case LLM_ARCH_MISTRAL4:
         case LLM_ARCH_MELLUM:
@@ -433,7 +436,7 @@ static bool arch_supported(const llm_arch arch) {
 
     // FIXME: these hit scheduler/view-backed-output issues with WebGPU on CI.
 #ifdef GGML_USE_WEBGPU
-    if (arch == LLM_ARCH_DEEPSEEK32 || arch == LLM_ARCH_GLM_DSA) {
+    if (arch == LLM_ARCH_DEEPSEEK32 || arch == LLM_ARCH_GLM_DSA || arch == LLM_ARCH_KIMI_K3) {
         return false;
     }
 #endif // GGML_USE_WEBGPU
diff --git a/tools/mtmd/CMakeLists.txt b/tools/mtmd/CMakeLists.txt
index 461c4f429cfb..b62c15c1570d 100644
--- a/tools/mtmd/CMakeLists.txt
+++ b/tools/mtmd/CMakeLists.txt
@@ -38,6 +38,7 @@ add_library(mtmd
             models/granite4-vision.cpp
             models/hunyuanvl.cpp
             models/internvl.cpp
+            models/kimik3.cpp
             models/inkling.cpp
             models/kimivl.cpp
             models/kimik25.cpp
diff --git a/tools/mtmd/clip-impl.h b/tools/mtmd/clip-impl.h
index fc011f714704..7844a6a81f96 100644
--- a/tools/mtmd/clip-impl.h
+++ b/tools/mtmd/clip-impl.h
@@ -41,6 +41,7 @@
 #define KEY_PROJ_DIM            "clip.%s.projection_dim"
 #define KEY_N_HEAD              "clip.%s.attention.head_count"
 #define KEY_N_HEAD_KV           "clip.%s.attention.head_count_kv"
+#define KEY_N_EMBD_HEAD         "clip.%s.attention.head_dim"
 #define KEY_LAYER_NORM_EPS      "clip.%s.attention.layer_norm_epsilon"
 #define KEY_FEATURE_LAYERS      "clip.%s.feature_layer"
 
@@ -405,6 +406,7 @@ enum projector_type {
     PROJECTOR_TYPE_YOUTUVL,
     PROJECTOR_TYPE_YASA2,
     PROJECTOR_TYPE_KIMIK25,
+    PROJECTOR_TYPE_KIMIK3,
     PROJECTOR_TYPE_NEMOTRON_V2_VL,
     PROJECTOR_TYPE_HUNYUANVL,
     PROJECTOR_TYPE_PARAKEET,
@@ -463,6 +465,7 @@ static std::map<projector_type, std::string> PROJECTOR_TYPE_NAMES = {
     { PROJECTOR_TYPE_YOUTUVL,           "youtuvl"},
     { PROJECTOR_TYPE_YASA2,             "yasa2"},
     { PROJECTOR_TYPE_KIMIK25,           "kimik25"},
+    { PROJECTOR_TYPE_KIMIK3,            "kimik3"},
     { PROJECTOR_TYPE_NEMOTRON_V2_VL,    "nemotron_v2_vl"},
     { PROJECTOR_TYPE_EXAONE4_5,         "exaone4_5"},
     { PROJECTOR_TYPE_HUNYUANVL,         "hunyuanvl"},
diff --git a/tools/mtmd/clip-model.h b/tools/mtmd/clip-model.h
index bfe021922098..ee6b09879b44 100644
--- a/tools/mtmd/clip-model.h
+++ b/tools/mtmd/clip-model.h
@@ -54,6 +54,8 @@ struct clip_hparams {
     int32_t projection_dim = 0;
     int32_t n_head = 0;
     int32_t n_head_kv = 0;
+    // 0 = derive from n_embd; set when qkv width != n_embd
+    int32_t n_embd_head = 0;
     int32_t n_layer = 0;
     int32_t n_merge = 1; // number of patch merges **per-side**
 
diff --git a/tools/mtmd/clip.cpp b/tools/mtmd/clip.cpp
index f1ec237a4bda..dd0c24b540c2 100644
--- a/tools/mtmd/clip.cpp
+++ b/tools/mtmd/clip.cpp
@@ -253,7 +253,7 @@ clip_graph::clip_graph(clip_ctx * ctx, const clip_image_f32 & img) :
         n_embd(hparams.n_embd),
         n_head(hparams.n_head),
         n_head_kv(hparams.n_head_kv),
-        d_head(n_head > 0 ? n_embd / n_head : 0),
+        d_head(hparams.n_embd_head > 0 ? hparams.n_embd_head : (n_head > 0 ? n_embd / n_head : 0)),
         n_layer(hparams.n_layer),
         n_mmproj_embd(clip_n_mmproj_embd(ctx)),
         eps(hparams.eps),
@@ -372,13 +372,13 @@ ggml_tensor * clip_graph::build_vit(
                 /* nb1    */ ggml_row_size(cur->type, d_head),
                 /* nb2    */ cur->nb[1],
                 /* nb3    */ cur->nb[1] * n_pos,
-                /* offset */ ggml_row_size(cur->type, n_embd));
+                /* offset */ ggml_row_size(cur->type, n_head * d_head));
 
                 Vcur = ggml_view_4d(ctx0, cur, d_head, n_head, n_pos, B,
                 /* nb1    */ ggml_row_size(cur->type, d_head),
                 /* nb2    */ cur->nb[1],
                 /* nb3    */ cur->nb[1] * n_pos,
-                /* offset */ ggml_row_size(cur->type, 2 * n_embd));
+                /* offset */ ggml_row_size(cur->type, 2 * n_head * d_head));
 
                 if (layer.q_norm) {
                     GGML_ASSERT(layer.q_norm->ne[0] == Qcur->ne[0]);
@@ -977,6 +977,10 @@ static std::unique_ptr<clip_graph> clip_get_graph_builder(clip_ctx * ctx, const
             {
                 builder = std::make_unique<clip_graph_kimik25>(ctx, img);
             } break;
+        case PROJECTOR_TYPE_KIMIK3:
+            {
+                builder = std::make_unique<clip_graph_kimik3>(ctx, img);
+            } break;
         case PROJECTOR_TYPE_COGVLM:
             {
                 builder = std::make_unique<clip_graph_cogvlm>(ctx, img);
@@ -1194,6 +1198,7 @@ struct clip_model_loader {
             const char * prefix = is_vision ? "vision" : "audio";
             get_u32(string_format(KEY_N_EMBD,         prefix), hparams.n_embd);
             get_u32(string_format(KEY_N_HEAD,         prefix), hparams.n_head);
+            get_u32(string_format(KEY_N_EMBD_HEAD,    prefix), hparams.n_embd_head, false);
             get_u32(string_format(KEY_N_FF,           prefix), hparams.n_ff);
             get_u32(string_format(KEY_N_BLOCK,        prefix), hparams.n_layer);
             get_u32(string_format(KEY_PROJ_DIM,       prefix), hparams.projection_dim);
@@ -1460,6 +1465,23 @@ struct clip_model_loader {
                         hparams.rope_theta = 10000.0f;
                         get_u32(KEY_PROJ_SCALE_FACTOR, hparams.n_merge, false);
 
+                        int min_pixels = 0, max_pixels = 0;
+                        get_u32(KEY_IMAGE_MIN_PIXELS, min_pixels, false);
+                        get_u32(KEY_IMAGE_MAX_PIXELS, max_pixels, false);
+                        if (min_pixels > 0 && max_pixels > 0) {
+                            hparams.image_min_pixels = min_pixels;
+                            hparams.image_max_pixels = max_pixels;
+                            hparams.warmup_image_size = static_cast<int>(std::sqrt(max_pixels));
+                        } else {
+                            hparams.set_limit_image_tokens(2, 4096);
+                        }
+                    } break;
+                case PROJECTOR_TYPE_KIMIK3:
+                    {
+                        hparams.image_resize_algo = RESIZE_ALGO_BILINEAR;
+                        hparams.rope_theta = 10000.0f;
+                        get_u32(KEY_PROJ_SCALE_FACTOR, hparams.n_merge, false);
+
                         int min_pixels = 0, max_pixels = 0;
                         get_u32(KEY_IMAGE_MIN_PIXELS, min_pixels, false);
                         get_u32(KEY_IMAGE_MAX_PIXELS, max_pixels, false);
@@ -2480,6 +2502,13 @@ struct clip_model_loader {
                     model.mm_2_w = get_tensor(string_format(TN_LLAVA_PROJ, 2, "weight"));
                     model.mm_2_b = get_tensor(string_format(TN_LLAVA_PROJ, 2, "bias"));
                 } break;
+            case PROJECTOR_TYPE_KIMIK3:
+                {
+                    // patchmergerv2, bias-free, norm after the projection
+                    model.mm_1_w = get_tensor(string_format(TN_LLAVA_PROJ, 1, "weight"));
+                    model.mm_2_w = get_tensor(string_format(TN_LLAVA_PROJ, 2, "weight"));
+                    model.mm_post_norm_w = get_tensor(string_format(TN_MM_POST_NORM, "weight"));
+                } break;
             case PROJECTOR_TYPE_KIMIVL:
             case PROJECTOR_TYPE_PADDLEOCR:
             case PROJECTOR_TYPE_KIMIK25:
@@ -3686,6 +3715,7 @@ int clip_n_output_tokens(const clip_ctx * ctx, const clip_image_f32 * img) {
         case PROJECTOR_TYPE_LFM2:
         case PROJECTOR_TYPE_KIMIVL:
         case PROJECTOR_TYPE_KIMIK25:
+        case PROJECTOR_TYPE_KIMIK3:
             {
                 // dynamic size
                 int out_patch_size = params.patch_size * ctx->model.hparams.n_merge;
@@ -4387,6 +4417,7 @@ bool clip_image_batch_encode(clip_ctx * ctx, int n_threads, const clip_image_f32
         case PROJECTOR_TYPE_PIXTRAL:
         case PROJECTOR_TYPE_KIMIVL:
         case PROJECTOR_TYPE_KIMIK25:
+        case PROJECTOR_TYPE_KIMIK3:
         case PROJECTOR_TYPE_LIGHTONOCR:
             {
                 // set the 2D positions
@@ -5078,6 +5109,7 @@ int clip_n_mmproj_embd(const struct clip_ctx * ctx) {
         case PROJECTOR_TYPE_KIMIVL:
         case PROJECTOR_TYPE_PADDLEOCR:
         case PROJECTOR_TYPE_KIMIK25:
+        case PROJECTOR_TYPE_KIMIK3:
         case PROJECTOR_TYPE_YASA2:
             return ctx->model.mm_2_w->ne[1];
         case PROJECTOR_TYPE_HUNYUANVL:
diff --git a/tools/mtmd/models/kimik3.cpp b/tools/mtmd/models/kimik3.cpp
new file mode 100644
index 000000000000..fe06c7fdbb38
--- /dev/null
+++ b/tools/mtmd/models/kimik3.cpp
@@ -0,0 +1,80 @@
+#include "models.h"
+
+#include <cmath>
+#include <cstring>
+
+// Kimi-K3 MoonViT-3d, image path.
+// Follows clip_graph_kimik25, but with RMSNorm, no biases, qkv width != n_embd, and a post-norm patchmergerv2 projector.
+// Images only: at t == 1 the temporal pool and the temporal position term vanish.
+
+ggml_tensor * clip_graph_kimik3::resize_position_embeddings_3d(uint32_t interpolation_mode) {
+    ggml_tensor * pos_embd = model.position_embeddings;
+    const int height       = img.ny() / patch_size;
+    const int width        = img.nx() / patch_size;
+
+    GGML_ASSERT(pos_embd);
+
+    const int64_t stored_c = pos_embd->ne[0];
+    const int64_t orig_w   = pos_embd->ne[1];
+    const int64_t orig_h   = pos_embd->ne[2];
+
+    GGML_ASSERT(stored_c == n_embd);
+
+    if (height == (int) orig_h && width == (int) orig_w) {
+        return ggml_cont_2d(ctx0, pos_embd, n_embd, width * height);
+    }
+
+    pos_embd = ggml_permute(ctx0, pos_embd, 2, 1, 0, 3);
+    pos_embd = ggml_interpolate(ctx0, pos_embd, height, width, n_embd, 1, interpolation_mode);
+    pos_embd = ggml_permute(ctx0, pos_embd, 2, 1, 0, 3);
+    pos_embd = ggml_cont_2d(ctx0, pos_embd, n_embd, width * height);
+    return pos_embd;
+}
+
+ggml_cgraph * clip_graph_kimik3::build() {
+    ggml_tensor * pos_h = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_patches);
+    ggml_set_name(pos_h, "pos_h");
+    ggml_set_input(pos_h);
+
+    ggml_tensor * pos_w = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_patches);
+    ggml_set_name(pos_w, "pos_w");
+    ggml_set_input(pos_w);
+
+    ggml_tensor * learned_pos_embd = resize_position_embeddings_3d(GGML_SCALE_MODE_BILINEAR);
+
+    // Q/K are de-interleaved during conversion.
+    auto add_pos = [&](ggml_tensor * cur, const clip_layer &) {
+        return build_rope_2d(ctx0, cur, pos_w, pos_h, hparams.rope_theta, false);
+    };
+
+    ggml_tensor * inp = build_inp();
+    inp = ggml_add(ctx0, inp, learned_pos_embd);
+
+    ggml_tensor * cur = build_vit(
+                            inp, n_patches,
+                            NORM_TYPE_RMS,
+                            hparams.ffn_op,
+                            nullptr,
+                            add_pos);
+    cb(cur, "vit_out", -1);
+
+    {
+        const int scale_factor = model.hparams.n_merge;
+        cur = build_patch_merge_permute(cur, scale_factor);
+
+        cur = build_ffn(cur,
+            model.mm_1_w, nullptr,
+            nullptr,      nullptr,
+            model.mm_2_w, nullptr,
+            FFN_GELU,
+            -1);
+        cb(cur, "proj_mlp_out", -1);
+
+        cur = build_norm(cur, model.mm_post_norm_w, nullptr, NORM_TYPE_RMS, hparams.eps, -1);
+        cb(cur, "proj_out", -1);
+    }
+
+    ggml_build_forward_expand(gf, cur);
+
+    return gf;
+}
diff --git a/tools/mtmd/models/models.h b/tools/mtmd/models/models.h
index 6735c9cdebf9..0a969dc56681 100644
--- a/tools/mtmd/models/models.h
+++ b/tools/mtmd/models/models.h
@@ -238,6 +238,13 @@ struct clip_graph_parakeet : clip_graph {
     ggml_cgraph * build() override;
 };
 
+struct clip_graph_kimik3 : clip_graph {
+    clip_graph_kimik3(clip_ctx * ctx, const clip_image_f32 & img) : clip_graph(ctx, img) {}
+    ggml_cgraph * build() override;
+
+    ggml_tensor * resize_position_embeddings_3d(uint32_t interpolation_mode);
+};
+
 struct clip_graph_exaone4_5 : clip_graph {
     clip_graph_exaone4_5(clip_ctx * ctx, const clip_image_f32 & img) : clip_graph(ctx, img) {}
     ggml_cgraph * build() override;
diff --git a/tools/mtmd/mtmd.cpp b/tools/mtmd/mtmd.cpp
index a70f326a9673..2f5fc9fdf099 100644
--- a/tools/mtmd/mtmd.cpp
+++ b/tools/mtmd/mtmd.cpp
@@ -567,6 +567,7 @@ struct mtmd_context {
                     image_preproc = std::make_unique<mtmd_image_preprocessor_dyn_size>(ctx_v);
                 } break;
             case PROJECTOR_TYPE_KIMIK25:
+            case PROJECTOR_TYPE_KIMIK3:
                 {
                     // GLM-5.2-V reuses the Kimi-K2.5 vision encoder and projector, but marks
                     // images with its own tokens, so decide based on the text model vocab