From a4005bd2c11c6e818eec8a5b4ea148768287404b Mon Sep 17 00:00:00 2001
From: Datta Nimmaturi <venkatadattasainimmaturi@gmail.com>
Date: Tue, 10 Feb 2026 06:31:36 +0000
Subject: [PATCH] Fix qwen3moe experts save

---
 convert_hf_to_gguf.py | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/convert_hf_to_gguf.py b/convert_hf_to_gguf.py
index 6e6e618989be..dce245f33833 100755
--- a/convert_hf_to_gguf.py
+++ b/convert_hf_to_gguf.py
@@ -4137,7 +4137,8 @@ def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iter
             # skip visual tensors
             return
         if name.find("experts") != -1:
-            n_experts = self.hparams["num_experts"]
+            # Qwen3MoE seems to use num_local_experts instead of num_experts
+            n_experts = self.hparams.get('num_experts', None) or self.hparams.get('num_local_experts')
             assert bid is not None
 
             if self._experts is None: