The best open-weight AI model ranks 12th, 30 points behind Claude Opus 5.5. The top six open models are all Chinese

The highest-ranked AI model with downloadable weights on ArtificialWatch's bench index is DeepSeek V4.1 Flash, in 12th place of 44 at 68. That is 30.1 points behind the leader, Anthropic's Claude Opus 5.5, at 98.1. All 11 models above it are closed: four from Anthropic, three from OpenAI, two from Google, one from Meta and one from Moonshot.

THE OPEN-WEIGHT TOP SIX ARE ALL CHINESE

- 12th: DeepSeek V4.1 Flash (DeepSeek), 68, MIT license. - 18th: MiMo-V2.6 Pro (Xiaomi), 49.1, MIT license. - 19th: GLM-5.3 (Zhipu AI), 48.1. - 26th: MiniMax M3.1 Flash (MiniMax), 38.5. - 29th: Step 5 Preview (StepFun), 34. - 31st: GLM-5.2 (Zhipu AI), 32.4, MIT license.

Ten of the 15 open-weight models on the index come from Chinese labs.

OUTSIDE CHINA

The best open-weight model from anywhere else is Mistral Large 4, from France's Mistral, in 34th place at 26.3. Its weights came out on October 29. The best from a US lab is Meta's Muse Spark Open, 39th at 15. The last three places on the index all go to US open-weight models: Reflection AI's Beam (42nd, 7.4) and Thinking Machines' Inkling (43rd, 3.8) and Inkling Small (44th).

ON CODING

The gap is smaller on ArtificialWatch's coding index, where DeepSeek V4.1 Flash is also the best open-weight model, 10th of 45 at 75.6, against Claude Opus 5.5's 96. Moonshot's Kimi K3 ranks far higher on coding (16th, 64.4) than on the main index (36th, 23.1).

PRICE

DeepSeek V4.1 Flash costs $0.30 per million input tokens and $1.20 per million output tokens on DeepSeek's API at weekday peak hours, and half that off-peak. Claude Opus 5.5 costs $4 / $20.

WHAT COULD CHANGE THIS WEEK

Moonshot's Kimi K3.1 Pro sits 11th at 69.9, 1.9 points above DeepSeek V4.1 Flash. Moonshot says its weights come by November 4. If they do, it becomes the top open-weight model on the index.

HOW THE INDEX WORKS

The bench index is the median of a model's percentile ranks across up to seven public leaderboards: LMArena, LiveBench, SimpleBench, LLM Stats, BridgeBench, Vals AI and Artificial Analysis. A model needs at least two of them to get a score. The table was refreshed on November 3. A model counts as open-weight when its weights are in a public Hugging Face repo; every repo was checked on October 29, and Muse Spark Open's requires approval to download. Gemini 4 Carbon, second at 96.2, is scored by public boards but is not yet on a public API. The full open-weights ranking, with size, license and repo for each model, is at artificialwatch.com/open-weights.