Mistral releases Large 4’s weights and tightens its license for API resellers

France’s Mistral AI released the weights of Mistral Large 4, its largest model, on October 29, 2026. Nicknamed “Le Chonk,” the model has about 1 trillion parameters, 49 billion of them active per token. The weights are on Hugging Face, and Mistral’s API charges $0.70 per million input tokens and $2.10 per million output tokens, or about €0.63 and €1.89.

The model ships under a new license, the Mistral Ecosystem Open-Weight License. It allows commercial use, modification and fine-tuning, with one exception: a company that sells access to the model as a service must first get a commercial license from Mistral. Its predecessor, Mistral Large 3, is licensed under Apache 2.0, which has no such clause.

Three more weeks of training after the preview

Mistral opened an API preview of Large 4 on October 6 and promised the weights by the end of the month. In between, the company says, it ran three more weeks of reinforcement learning. According to Mistral, the model’s score on Terminal-Bench 4.0, a test of command-line tasks, rose from 28% to 41%, with 30% fewer output tokens.

Mistral also reports 19.7% on FrontierCode 2.0, a private benchmark from Cognition, ahead of the other open models it compared. On Artificial Analysis’s Coding Agent Index, Large 4 scores 48%, effectively tied with China’s GLM-5.3 (49%), while using 47% fewer tokens per task on average. In Surge AI’s blind code-quality evaluation, the only model preferred over it was Anthropic’s Claude Opus 5.5, which won 78% of their head-to-head comparisons.

Mistral says it trained the model from scratch on 3,800 NVIDIA GPUs in its own data centers in Europe.

A license aimed at API providers

The Large 4 license is the MIT license plus one clause. If the licensee or one of its affiliates runs a “Model as a Service” business, it must enter into a commercial license with Mistral before any commercial use of the model or its derivatives. The license defines that as giving a third party access to the model or its derivatives for inference or post-training, by any means (an API or a hosted endpoint, for example), in a way that lets the third party exercise meaningful control over inputs, parameters or training data.

Two cases fall outside that definition: an end-user product that embeds the model only in specific features, and simply relaying requests to a model someone else hosts. Internal use is also exempt, as is any use through Mistral’s official products.

Pricing against OpenAI and Anthropic

The table compares Mistral’s price, in effect since October 29, with the published standard rates of OpenAI and Anthropic, excluding tax. Euro figures use the European Central Bank’s October 29 reference rate (€1 = $1.1130).

Price per token is not cost per task, which depends on how many tokens a model burns through. When OpenAI launched GPT-6.1 Terra on October 26, it pitched the model on cost per task against Claude Opus 5.5.

The model is also available on OpenRouter, through NVIDIA and on Microsoft Foundry.

The hardware it takes to self-host

Self-hosting the full versions, the approach known as local AI for business, requires data center hardware. The FP8 version of Large 4 is about 1.06 TB and the compressed NVFP4 version about 610 GB, according to the vLLM project’s deployment recipe. vLLM validated the NVFP4 version on eight-GPU B200 or B300 servers and on four GB300 GPUs. It also validated the FP8 version on two eight-GPU H200 servers.

In its launch post, Hugging Face says the 2-bit quantized versions fit on a single Mac with an Ultra-series chip and 512 GB of unified memory, running at about 15 tokens per second on an M3 Ultra. The model card recommends Mistral’s API if local serving falls short.

Banks and cyber defenders among the partners

Large 4 scores 51% on Artificial Analysis’s Cyber Index, which measures how well a model finds, reproduces and fixes flaws in code. On the chart Mistral published, two Chinese open models do better: DeepSeek V4.1 Flash 1017 (57%) and Xiaomi’s MiMo-V2.6-Pro (56%). Mistral says Large 4 refuses 96.6% of malicious cyber requests on average across three public test sets.

A specialized variant for penetration testing and red teaming is available only to vetted partners such as Orange Cyberdefense, under contract. Testing is limited to systems those partners own or are explicitly authorized to test. Google made a similar call on October 28, giving a subset of cyber defense specialists first access to Gemini 4 Carbon without releasing its weights.

About 20 organizations are quoted in the announcement. Belgian bank Belfius says it can fine-tune the model, host it on its own infrastructure and integrate it into its digital channels. France’s La Banque Postale says that for a public bank, technological sovereignty is a key prerequisite. ENISA, the European Union’s cybersecurity agency, is testing Large 4 to find vulnerabilities, audit code and analyze security logs.

Pierre Stock, Mistral’s vice president of science, summed up the pitch to TechCrunch: businesses and governments get frontier AI to defend themselves, with “access that can’t be cut off by definition because the model is open-weighted.” Writing for the French tech site Next, Vincent Hermann said the model should put the company back in the race “after long months of near-silence on large models” (our translation).

The limits of the published numbers

Mistral picked the benchmarks it highlights. The preview score and the refusal rate are its own measurements. The 48% and 41% come from Artificial Analysis’s Coding Agent Index, which ran Large 4 inside Mistral Vibe, Mistral’s own coding agent. Mistral has promised a full technical report but given no date.

Artificial Analysis’s Intelligence Index (version 5.0, nine evaluations), checked on October 30, is less flattering. Large 4 scores 26 there, versus 32 for DeepSeek V4.1 Flash 1017 and GLM-5.3, 36 for Claude Haiku 5.5, 51 for GPT-6.1 Terra and 58 for Claude Opus 5.5. The index includes Terminal-Bench 4.0 and FrontierCode 2.0, two tests where Mistral claims the lead among open models. On those, Large 4 scores 24.7% and 15%, behind GLM-5.3 (41.9% and 21.6%) and DeepSeek V4.1 Flash 1017 (33.8% and 28.5%).

The model card acknowledges that the best Chinese open models score higher on this index and calls Large 4 the top-ranked Western open model. The Artificial Analysis table bears that out: six Chinese open models rank ahead of it, while Meta’s Muse Spark Open scores 25 and NVIDIA’s Nemotron 3.5 Super VL, released October 15, scores 14.

Mistral also flags a risk. Monitoring cybersecurity use since the preview, the company saw the model refuse black-box attacks on live targets, but it says the model could still help carry out real attacks. Mistral says it is changing its guardrails and plans to keep updating them in the coming months.

What it means

Large 4 is the highest-ranked open-weight model built outside China on Artificial Analysis’s index, and it was trained in Europe. The most obvious users are banks, public agencies and security teams that need to keep both data and model under their own control and have multi-GPU servers to run it. For everyone else, the simplest route is Mistral’s API or a cloud such as Microsoft Foundry, at a per-token price below that of GPT-6.1 Terra and Claude Sonnet 5.5.

The coding gains Mistral announced rest on benchmarks it chose and, for the 41% on Terminal-Bench, on its own agent. Artificial Analysis’s Intelligence Index puts Large 4 far behind the best closed models and behind six Chinese open models. The open question on the license is how Mistral will apply “meaningful control,” the test that separates free use from resale under contract.

Three things to watch: the promised technical report, Large 4 results with coding agents other than Mistral Vibe, and how the guardrail changes apply to copies already downloaded, which Mistral does not say. The company also plans a series of specialized models built on Large 4, with no timeline.

Sources

• Mistral Large 4 (Le Chonk) is now open weight, Mistral AI, October 29, 2026
• Introducing Mistral Large 4, Mistral AI, October 6, 2026
• Mistral Large 4 – Le Chonk (model card), Mistral AI on Hugging Face, accessed October 30, 2026
• Mistral Ecosystem Open-Weight License, Mistral AI on Hugging Face, accessed October 30, 2026
• Artificial Analysis Intelligence Index v5.0, Artificial Analysis, accessed October 30, 2026
• Mistral Large 4 recipe, vLLM, accessed October 30, 2026
• Mistral releases weights for its trillion-parameter AI model, targeting cybersecurity, TechCrunch, October 29, 2026
• Mistral lance la version finale de Large 4 et publie ses poids ouverts, Next, October 29, 2026

Header image: MACCUS chart, data from Artificial Analysis (Intelligence Index v5.0, accessed October 30, 2026).