Mistral Large 4 Open Weights: Fast Coder, Not a Frontier Model

Mistral calls Large 4 the best open agentic coding model in the world, and on Terminal-Bench 4.0 it leads GLM-5.3 by one point, 41% to 40%. On Artificial Analysis’ broader Intelligence Index it scores 26, six points behind the best Chinese open models. For teams choosing an open-weight model to self-host for coding agents, the useful numbers are elsewhere: speed, tokens per task and the license.

What Happened

The Decoder reports that Mistral has put Large 4, nicknamed “Le Chonk”, on Hugging Face, three weeks after a public preview. It has about one trillion parameters with 52 billion active per step, takes text and images in, and outputs text and code. SGLang’s docs say the standard version runs on eight Nvidia B300 GPUs, and a more compressed one on four GB300s.

Reinforcement learning since the preview lifted its Terminal-Bench 4.0 score from 28% to 41%. Mistral says it reaches 48 on Artificial Analysis’ Coding Agent Index, practically tied with GLM-5.3 while using 47% fewer tokens. Claude Code with Opus 5.5 tops that index at 62. On Cognition’s FrontierCode 2.0, Large 4 leads open models at 19.7%, against 46% for Opus 5.5.

Why It Matters

The independent picture is more mixed than Mistral’s. Artificial Analysis rates the final model at 26 on its Intelligence Index, one point above the preview and above the 23-point median for open models its size, but behind DeepSeek V4.1 Flash and GLM-5.3 at 32. Mistral’s lead may be brief: Moonshot plans to release Kimi K3.1 Pro’s weights by November 4, and it scores 67% on Terminal-Bench 4.0.

On cyber defense, Mistral calls Large 4 the leading open model, but its own chart puts it third among open models on Artificial Analysis’ Cyber Index at 51%. It does refuse 96.6% of malicious cyber requests, more than any other open model, Mistral says.

Speed and cost are the clearer wins. Large 4 generates 161 tokens per second, and an index task costs $2.06 on average, down from $5.54 for the preview. That’s cheaper than GLM-5.3 and Kimi K3.1 Pro, but far more than DeepSeek V4.1 Flash at $0.60. API pricing is $0.70 per million input tokens and $2.10 per million output.

What This Means for Builders

The license matters for platform teams. It’s a modified MIT license: use, modification and commercial deployment are free, but offering Large 4 as a commercial cloud service needs a separate license. The New Stack quotes Mistral VP of Science Pierre Stock on why open weights matter: “nobody can switch it off because it can run either on your infrastructure or on an infrastructure that you can trust.” Mistral pitches the model as sovereign AI, trained entirely in Europe.

That argument is strongest for teams with data-residency or vendor-risk requirements, where Stack Archive has covered Anthropic’s EU sovereignty push. For pure cost per coding task, the DeepSeek comparison is hard to ignore, as our look at DeepSeek V4.1 Flash’s agentic costs showed.

Source: The Decoder

Content created with AI assistance and reviewed for accuracy.