1.05T Parameters, Only 52B Activated: Mistral Large 4 Leads by Just 1 Point After Open-Weight Release

Mistral AI has released the weights for its flagship model, Mistral Large 4. The Artificial Analysis evaluation results are straightforward: the High reasoning configuration scores 26 on Intelligence Index v5, 1 point more than Preview's 25, ranking first among open-weight models not developed in China and leading Muse Spark Open (xhigh) by only 1 point.

First, the model specs: 1.05T total parameters, 52B activated parameters, a 1M-token context window, and support for text and image input with text output. Pricing is $0.70/$2.10 per million input/output tokens, and $0.07 for cached input. The license is the MEOW License (Mistral Ecosystem Open-Weight License); operating a Model-as-a-Service business requires negotiating a separate commercial license with Mistral. This detail affects how most teams can use it more than benchmark scores do.

On rankings, Mistral Large 4 (high) scores 51 on the Cyber Index, ranking third among open-weight models and trailing DeepSeek V4.1 Flash 1017 (max, 57) and MiMo-V2.6-Pro (56). On the Intelligence Index, it trails DeepSeek V4.1 Flash 1017 (max) and GLM-5.3 (max) by 6 points. The gap is mainly in coding and agentic knowledge work: FrontierCode 2.0 is only 15%, versus 29% for DeepSeek V4.1 Flash 1017 (max) and 22% for GLM-5.3 (max); AA-Briefcase Elo is 1349, versus 1570 for DeepSeek and 1504 for GLM.

Efficiency is the focus of this update. Mistral Large 4 (high) uses 82K output tokens per Intelligence Index task, 39% fewer than Preview's 135K, putting it on the open-weight Pareto frontier. The seven open-weight models with higher scores use 113K to 225K output tokens per task. It does not raise the score by much, but it brings down how many tokens must be burned to complete an evaluation run.

On the Coding Agent Index, Mistral Large 4 (high) scores 48 in Mistral Vibe, using 8.5M tokens per task. GLM-5.3 (max) scores 49 in Opencode, using 15.9M tokens; DeepSeek V4.1 Flash 1017 (max) scores 46 in DeepSeek Harness, using 16.7M tokens. It uses roughly half the tokens, with its score falling between the two.

Hallucination rate also dropped. In the updated AA-Omniscience v1.1, Mistral Large 4 (high) has a 28% hallucination rate, down from Preview's 41% and level with Muse Spark Open (xhigh). But accuracy is only 26%, close to Ling 3.1 Flash 1019's 27% and below DeepSeek V4.1 Flash 1017 (max)'s 40% and Muse Spark Open (xhigh)'s 42%. Low hallucination does not equal high accuracy; these two metrics need to be viewed separately.

On cost, the new pricing is about 50% lower than Preview. At the new prices, each Intelligence Index task costs $2.06, down 63% from Preview's $5.54. That is close to Muse Spark Open (xhigh, $2.02), but higher than DeepSeek V4.1 Flash 1017 (max, $0.60) and MiMo-V2.6-Pro ($0.26).

The deployment barrier deserves separate mention. 1.05T total parameters and 52B activated parameters do not mean you only need to fit 52B on a machine. The full weights still have to be stored; activated parameters only determine which parts participate in computation during inference. Some commenters noted that a 1-bit Unsloth GGUF can already run on a Mac Studio, while others joked that loading a trillion parameters into memory would defeat many homelabs. Open weights and comfortable local running are two different things.

The license also drew discussion. Some commenters reminded that operating the model as a hosted service under the MEOW license requires a separate agreement with Mistral, so for most teams the first-party API may remain the main way to use it. Others joked that open weights come from everyone except the company with Open in its name, and that Mistral named the license MEOW and just kept going.

Full results and more details are available on Artificial Analysis's model page: https://artificialanalysis.ai/models/mistral-large-4