Back to LLM Models

Mistral releases Le Chonk's open weights and halves the API price

Mistral released the weights of Mistral Large 4 (Le Chonk) on October 29, three weeks after the public preview and with another reinforcement learning run on top of it. The same announcement cut the API price almost in half, from $1.36 per million input tokens and $4.18 per million output tokens to $0.70 and $2.10, with cached input at $0.07. The model is a 1 trillion parameter natively multimodal MoE with 52 billion active parameters, and the weights are now on Hugging Face.

The license: MIT, with one condition

The weights ship under the Mistral Ecosystem Open-Weight license, a modified MIT. Per the company's announcement, the model is free to use, modify, fine-tune and commercialize, with a single condition: a business that serves the model as a Model-as-a-Service offering has to buy a separate commercial license from Mistral. If you self-host it for your own workloads, the condition does not apply. The model is also available through the Mistral API, OpenRouter, Microsoft Foundry and NVIDIA NIM containers, for teams that want the model without the infrastructure.

What three weeks of RL changed

The coding numbers moved. On the Artificial Analysis Coding Agent Index, an independent evaluation of agentic coding, Mistral Large 4 scores 48, effectively tied with GLM-5.3 at the top of the open-weights class, while using 47% fewer tokens per task on average, according to Mistral. On Terminal-Bench 4.0 the score went from 28% in the preview to 41%, with output tokens down 30%. Cognition's FrontierCode 2.0 is a private benchmark, but Mistral reports a lead of 19.7% on it; it grades whether a repository maintainer would merge the agent's change. In a blind side-by-side evaluation run by Surge, graders preferred Mistral Large 4 over GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash, behind only Claude Opus 5.5.

The model also now offers three reasoning modes, Low, Medium and High (the default), so you can trade cost against precision per task instead of taking one fixed setting.

The cyber story, and the one caveat

Defensive security was the preview's strongest thread, and it remains the story. On the Artificial Analysis Cyber Index, which grades a model's ability to find, reproduce and patch vulnerabilities in a codebase, Mistral Large 4 leads at 51%. On the jailbreak battery the model refuses 96.6% of malicious cyber requests on average, the highest rate Mistral says it measures among open models.

There is a caveat, and Mistral states it plainly. While monitoring live traffic, the model refused to carry out black-box attacks against live targets, but the company says it saw a potential for uplift to conduct real attacks, and it is modifying the guardrails over the coming months. It is also giving a specialized, cybersecurity-optimized variant of the model to vetted enterprise partners, with testing restricted to owned or explicitly authorized systems.

If you run it yourself

The self-deployment math is now real. The model downloads from Hugging Face and serves with vLLM or SGLang, and the MaaS clause in the license means the only way to put it behind your own public API is a separate commercial deal. If you do not want the capex, the halved API pricing and the $0.07 cached-input rate are the numbers to compare against your current provider. The strongest published results are in agentic coding and defensive security, so point your test traffic there before you route anything in production.

Source: mistral.ai

More news

Back to LLM Models

Have something in mind?

Start a conversation