Skip to content

fix(mpc): Beam merges, Ultrafast states 1000 tps, AA index version surfaced - #15

Merged
JeffreyJYZ merged 1 commit into
mainfrom
fix/beam-ultrafast
Nov 9, 2026
Merged

JeffreyJYZ merged 1 commit into
mainfrom
fix/beam-ultrafast

Conversation

@jyzbot

@jyzbot jyzbot commented Nov 9, 2026

Copy link
Copy Markdown
Collaborator

Three things from the pasted table.

Beam is one model, not two

Beam Free (OpenCode) and Beam 1.0 (CommandCode) are the same model. The trailing Free is stripped generically, so the bare beam key now aliases to beam10:

before  Beam 1.0   CC only, free     |  Beam Free  OC only, free
after   Beam 1.0   free both sides, ability 17.7, tie

Matched models 26 → 27, only-in-OpenCode 7 → 6, unscored rows 19 → 16.

Ultrafast's speed

DeepSeek V4.1 Flash Ultrafast had no throughput. The suffix rule added ultrafast (longest-first, so fast cannot leave a dangling ultra), which would derive 512 from its base — but CommandCode publishes 1000. TPS_OVERRIDES carries the stated figure and lookupTps prefers it over anything derived:

DeepSeek V4.1 Flash Ultrafast   ability 39.5   tps 1000

AA's index version

Straight from their API docs:

artificial_analysis_intelligence_index is a composite score… Current version: v5.0.1. … Field names use snake_case. We keep the wire format stable: fields are deprecated, not renamed.

So mpc has been reading v5 scores all along — but it discarded intelligence_index_version, which is the only thing that makes a v4/v5 mix legible. The footer label now prints it:

ability   Artificial Analysis (120 models, 422 speed, index v5.0.1)

Quota must not break every run

The free tier is 100 requests per 24 h, and the probe work spent today's. --bench aa now falls back to the keyless page on 429/401/410 and says so, instead of dying:

ability   Artificial Analysis (web, partial — API: ... 429: daily quota spent)

Verified with a real exhausted quota: the table still fills 38 values from CommandCode plus whatever the page carries.

Gates

bun check         No type errors in 89 files
bun test          115 pass, 0 fail  (+1 index-version test, +1 Beam alias assertion)
bun fmt --check   All 95 files are formatted
biome lint        clean

Version 0.1.10.

…a handled

Beam Free (OpenCode) and Beam 1.0 (CommandCode) are the same model; the bare key
now aliases to beam10, so one row replaces two free halves of a tie.

DeepSeek V4.1 Flash Ultrafast had no speed: the suffix rule would have derived 512
from its base, but CommandCode publishes 1000, so TPS_OVERRIDES carries the stated
figure and lookupTps prefers it over anything derived.

AA versions its index in-band -- v5.0.1 since 2026-10-27 -- while the field name
stays artificial_analysis_intelligence_index. The footer label now prints that
version, because a v4/v5 mix otherwise reads as models moving. A 429 (spent daily
quota), 401 or 410 now falls back to the keyless page and says so, instead of
failing every run.

Version 0.1.10.
@vercel

vercel Bot commented Nov 9, 2026

Copy link
Copy Markdown

Deployment failed for project cmdcode-tools-site with the following error:

Resource is limited - try again in 24 hours (more than 100, code: "api-deployments-free-per-day").

Learn More: https://vercel.com/jeffreys-projects-b51c8adf?upgradeToPro=build-rate-limit

@JeffreyJYZ
JeffreyJYZ merged commit 6b96599 into main Nov 9, 2026
8 of 9 checks passed
@JeffreyJYZ
JeffreyJYZ deleted the fix/beam-ultrafast branch November 9, 2026 14:39
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants