Mistral AI Drops 'Le Chonk': A Massive AI Model Named After a Cat Meme

What it claims

Mistral Large 4 is a 1-trillion-parameter mixture-of-experts model with 49B active parameters per query. It was launched Oct. 6 at $1.36 per million input tokens and $4.18 per million output tokens. Chief scientist Guillaume Lample calls it "by far the strongest open-weight model from the US or Europe." The scores below place it at or near the top of the open-weight field Mistral chose to compare against. It trails the closed frontier models on most of the benchmarks reported.

Evidence

The article quotes Mistral's launch chart and some third-party leaderboards. It does not give evaluation settings.

About AutomationBench: it has 657 tasks in simulated business software covering finance, HR, sales and support. It scores the share of each task's goals completed and gives zero credit if the model breaks a rule.

The pattern is that Large 4 is competitive with the Chinese open-weight models (DeepSeek V4 Pro, Kimi K3, GLM-5.3, Qwen3.8 Max). It is roughly 7–18 points behind the best closed models on the coding and agentic benchmarks. Finance Agent v2 is the one listed case where it beats a closed frontier model.

How to use it

Price against closed models, per million tokens:

That makes Large 4 about 1/3 of Opus 5.5's price on input and 1/5 on output. Against Astra it is about 1/7 on input and 1/12 on output.

Scale against the previous flagship: Large 3 had 41B active parameters out of 675B. Large 4 raises this to 49B active out of 1T. Per-query compute grows much less than total size, but self-hosting still means holding all 1T parameters.

Weights: Mistral promises them "by the end of October." The article gives no licence, context length, API model identifier or hardware guidance.

Limits

• Weights are not yet out. The open-weight claim and the benchmarks cannot be verified independently until they are. The licence is unstated, so check it before planning self-hosting or commercial use.
• The comparisons were chosen by Mistral. Its announcement benchmarks mainly against Chinese open-weight models. Claude and GPT appear in only a few comparisons.
• The article contradicts itself about Opus 5.5. It says Opus 5.5 appears "only in a cybersecurity claim," yet it also cites an Opus 5.5 score of 58.6 on Finance Agent v2 from the launch chart.
• Some numbers come from other sources. The AutomationBench comparisons come from Artificial Analysis, and the 74 DeepSWE scores come from Datacurve's leaderboard. These may not match the conditions behind Mistral's own scores.
• This is a news article, not the launch post. It gives no details on prompting, evaluation harness or sampling settings.