What Mistral released
Mistral AI, the Paris-based company that has become the most visible European challenger to OpenAI, Anthropic and Google, has put the Large 4 weights online. The nickname “Le Chonk” is doing a lot of work: this is not a small model. It has roughly one trillion parameters in total, with 52 billion active per step, and it accepts both text and images while producing text and code.
The hardware requirement settles the self-hosting question quickly. According to the docs for the inference framework SGLang, the standard version runs on eight Nvidia B300 GPUs, while a more heavily compressed variant runs on four GB300 GPUs. That makes Le Chonk a datacentre decision, not a 16 GB workstation one. Quantised 7B–14B models fit a 16 GB card; Le Chonk does not.
Since the public preview three weeks earlier, Mistral kept training the model with reinforcement learning. The company says this brought big gains, especially in coding and cyber defense. Mistral researchers Pierre Stock and Corentin Petit walked through the process on a press call: the team built tasks in-house that got harder as training went on, and they said the only limit on performance is how difficult a task the team can still create and verify. Mistral calls it a “two-year bet” that is now paying off.
What the coding benchmarks say
After three more weeks of training, Large 4’s score on Terminal-Bench 4.0 rose from 28 to 41 percent. That puts it just ahead of GLM-5.3 at 40 percent. But Kimi K3.1 Pro tops that benchmark at 67 percent.
Mistral says Large 4 hits 48 percent on Artificial Analysis’ Coding Agent Index. That is practically tied with GLM-5.3, but Large 4 uses 47 percent fewer tokens to get there. Claude Code with Claude Opus 5.5 tops the index at 62 points.
On Cognition’s new FrontierCode 2.0 benchmark, Large 4 leads the open models at 19.7 percent. Claude Opus 5.5, however, scores 46 percent. In blind comparisons run by Surge, testers preferred Large 4’s code over code from GLM-5.3, Kimi K3 and DeepSeek more often than not. Against Opus 5.5, Large 4 won only 22 percent of matchups.
Mistral’s comparisons leave out Moonshot’s Kimi K3.1 Pro. Technically that is fair, since its weights are not public yet; Moonshot plans to release them by November 4, 2026. After that, Mistral’s lead may not last, because Kimi K3.1 Pro already tops Terminal-Bench 4.0 at 67 percent.
Where it sits next to Chinese models
Artificial Analysis’ Intelligence Index v5 gives the final version of Large 4 26 points in high mode. DeepSeek V4.1 Flash and GLM-5.3 each score 32 points, and MiMo-V2.6-Pro gets 30. The median for open models of similar size is 23 points, so Mistral beats the average — but not the strongest Chinese open models.
The preview scored 25 points, so the three extra weeks of training added just one point overall. At the very top of the index, Google’s Gemini 4 Carbon leads with 59 points, one point ahead of Claude Opus 5.5 at 58. The top open model is Kimi K3.1 Pro at 46 points, but its weights have not been released yet.
For European teams, that gap is the practical buying question. A model can be good enough for a coding assistant and still lose on the overall index to cheaper or stronger open-weight alternatives from China. The licence and the deployment location then decide whether those alternatives are even allowed inside the company.
Cyber defense: third among open models, not first
Mistral calls Large 4 the leading open model for cyber defense. The company’s own chart tells a different story. On the Artificial Analysis Cyber Index, Large 4 scores 51, behind DeepSeek V4.1 Flash at 57 and MiMo-V2.6-Pro at 56. That puts Large 4 third among open models, just ahead of GLM-5.3 Flash at 50. Kimi K3.1 Pro leads the entire index at 64.
The extra training barely moved this number. According to Artificial Analysis, the preview scored 50 points, and the final version is only one point higher. Large 4 does better at refusing harmful requests: Mistral says it rejects 96.6 percent of malicious cyber requests on average, more than any other open model. It is still supposed to handle legitimate defensive tasks. Mistral admits the model could make real attacks easier, though, so it plans to tighten safeguards in the coming months. Select partners like Orange Cyberdefense are already testing a special version for penetration testing; access requires a review and comes with contractual restrictions.
Price, licence and speed
Through the API, Large 4 costs $0.70 per million input tokens and $2.10 per million output tokens. During the preview, the list price was $1.36 and $4.18. Mistral cut it to $0.68 and $2.09 for two weeks at launch, so the new price roughly matches that old discount. For comparison, Kimi K3.1 Pro costs $2 for input and $10 for output.
Per Artificial Analysis, an index task costs $2.06 on average with Large 4, down from $5.54 for the preview. That makes Large 4 cheaper than GLM-5.3 at $5.01 and Kimi K3.1 Pro at $3.68, but much more expensive than DeepSeek V4.1 Flash at $0.60 and MiMo-V2.6-Pro at $0.26. Compared with other open models of similar size, Artificial Analysis still considers its per-token price on the high side.
Speed is a genuine strength. Large 4 generates 161 tokens per second, up from 111 for the preview. In Artificial Analysis’ head-to-head comparison, only DeepSeek V4.1 Flash and Muse Wonder are faster. On token efficiency, the model generated 110 million output tokens across the full index, compared with 150 million for the preview; the median for comparable models is 100 million, so Large 4 still sits slightly above it.
The weights are released under a modified MIT license. Use, modification and commercial deployment are free. Companies that offer Large 4 as a commercial cloud service, however, need a separate license from Mistral. The company mainly pitches the model as “sovereign” AI, since it was trained entirely in Europe and can run on your own servers. If you do not want to host it yourself, it is available on Microsoft Foundry, among other platforms.
The European angle
Mistral is a French company, and for a European buyer that changes the paperwork more than the benchmark. A processor agreement, a stated inference region and a vendor that answers to EU data protection law are easier to obtain from a Paris-based supplier than from a US hyperscaler — and easier still when the model runs on your own servers.
The AI Act adds a second layer. Since 2 August 2025, providers of general-purpose AI models placed on the EU market have had to publish technical documentation, a summary of the training data and a copyright policy. For a European deployer, the practical question is whether the vendor’s documentation is good enough for a compliance file. Mistral publishes licence terms and a model card; that is the material a procurement review will ask for, and it is one reason European companies keep a Mistral option on the shortlist even when a Chinese open-weight model scores higher.
None of this makes Le Chonk the better model. It makes it the easier one to defend in a procurement review. That is a different question, and sometimes it is the one that decides the purchase.
How to test it on your own code
Public tables rarely match a specific repository. Take twenty tasks from your own backlog, drawn from real pull requests rather than puzzles, and run each model through the same agent. Count how often the patch passes your test suite without human edits. That number is the only one your team should trust.
Watch three costs at once: attempts per merged change, wall-clock latency, and tokens burned. A coding agent that reads a large repository on every task spends most of its budget on input tokens, so a model that is cheap per token and needs three tries can lose to one that costs more and succeeds first time. Compare cost per finished task, not cost per million tokens.
Local running is the other variable. We test models on a 16 GB card in AI Arena, and VRAM decides what is testable at all before speed or quality enter the picture. Quantised 7B–14B models fit a 16 GB card, but Le Chonk does not. Per the SGLang docs, the standard version needs eight Nvidia B300 GPUs, and a heavily compressed variant still needs four GB300 GPUs. Self-hosting Le Chonk is therefore a datacentre decision, not a workstation one.
Can I run Le Chonk on a workstation?
No. The standard version runs on eight Nvidia B300 GPUs, and a heavily compressed variant runs on four GB300 GPUs, according to the SGLang docs. Quantised 7B–14B models fit a 16 GB card; Le Chonk does not.
Is the licence really MIT?
It is a modified MIT license. Use, modification and commercial deployment are free, but companies that offer Large 4 as a commercial cloud service need a separate license from Mistral.
Does the AI Act apply if I only use the model?
Using a general-purpose model as a company makes you a deployer rather than a provider, and the duties are lighter. They change if you fine-tune it and place it on the market under your own name, which is the point where documentation and transparency obligations land on you.
