Mistral Large 4 Le Chonk Open Weights: Europe's Trillion-Parameter AI Model Anyone Can Now Download

Europe's trillion-parameter AI model is now free to download. Here is what that means.

Mistral AI has opened the weights of Mistral Large 4, its 1-trillion-parameter flagship nicknamed Le Chonk, which means any business, government or developer with the right hardware can now download one of the world's most capable open models and run it on its own machines. The French lab released the weights on Hugging Face after three additional weeks of reinforcement learning that pushed the model to the front of the open-weight field in coding and defensive cybersecurity, and the release is a genuine test of whether open models can now sit at the same table as the closed systems sold by the big US labs.

What exactly did Mistral release?

Mistral Large 4 is a natively multimodal mixture-of-experts model with roughly 1 trillion total parameters, of which about 50 billion are active on any given token, and a context window that stretches to one million tokens. It entered public preview on October 6, 2026, and the open-weight release followed on October 29 after the extra training pass. Mistral says the model was trained from scratch on European infrastructure, which reporting puts at around 3,800 NVIDIA Grace Blackwell GPUs in the company's own European data centres, and the company frames that lineage as the foundation for sovereign AI: a model whose history from training to release is documented and whose weights belong to whoever downloads them.

How good is it at coding and cybersecurity?

Mistral reports that Large 4 leads open-weight models on Cognition's FrontierCode 2.0 benchmark with a score of 19.7 percent, and on Terminal-Bench 4.0 it improved 13 percentage points over its preview to 41 percent while cutting output tokens by 30 percent. On the Artificial Analysis Coding Agent Index it scores 48 percent, effectively tied with the leading open-weight rival GLM-5.3 while using 47 percent fewer tokens per task on average. In Surge Code's blind human evaluations, graders preferred it over GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash, ranking it behind only Claude Opus 5.5 among the models tested.

The cybersecurity story is the more unusual one. Mistral says the model scores 51 percent on the Artificial Analysis Cyber Index, which measures a model's ability to find vulnerabilities in a codebase, reproduce and validate them, and it reports a 96.6 percent average refusal rate on malicious cybersecurity requests, the highest it measured among open models. That combination, strong defensive security with strong refusals of offensive use, is why the model is marketed as built for defenders and why partners such as Credit Agricole's AI unit have highlighted it for security work.

What does "open weight" actually mean?

Open weight is not the same as open source. The trained weights are freely downloadable and can be used commercially, but Mistral has not released the training data or its internal tooling, and commercial model-as-a-service providers need a separate licence. Organizations planning to redistribute the weights or deploy them in regulated, high-risk settings should read the licence on the Hugging Face model card before committing, because the fine print governs what is allowed.

What does it cost to run?

The model now offers three reasoning modes, Low for simple tasks, Medium for everyday work and High as the default for the hardest cases, which lets developers trade quality against token spend. Reports put Mistral's API pricing at 0.70 dollars per million input tokens and 2.10 dollars per million output tokens. Self-hosting is a different matter: although only a slice of the parameters computes at any moment, all trillion of them must sit in accelerator memory, so running it in-house takes server-class, multi-GPU hardware rather than a workstation, with quantized variants trimming the bill.

Why this matters for businesses and governments

For companies and public institutions outside the United States and China, the release matters for a simple reason: it puts a frontier-class model inside the reach of any organization that prefers to keep its data and its AI stack under its own control. A Nigerian bank, telecom or government agency that cannot send sensitive data to a foreign API now has a credible option to run serious AI in-house, and Mistral's emphasis on European training infrastructure speaks directly to the sovereignty argument. For the wider market, a strong open release also keeps pressure on the closed vendors, which is part of why pricing across the industry has started to cool.

Frequently asked questions

Is Le Chonk really open source?

No. It is open weight: the parameters are downloadable and commercially usable under the licence, but the training data and tooling stay with Mistral.

Can it run on a laptop?

No. The full model needs multi-GPU server hardware even in quantized form, though smaller distilled variants are the practical path for routine tasks.

How does it compare with closed frontier models?

Mistral claims the lead among open-weight models in coding and says its visual grounding results beat GPT-6 Astra on the Dense 200 benchmark by a point, while remaining behind the very best closed systems on several measures.

Sources: Mistral AI announcement, "Mistral Large 4 (Le Chonk) is now open weight" (mistral.ai); Hugging Face model card mistralai/Mistral-Large-4-1T-A52B; Cryptonomist, November 3, 2026; AI Insider, November 2, 2026. Related reading on this blog: OpenAI's $11 billion raise and cooling AI prices and Anthropic's $3.4 billion power deal.