Mistral’s Le Chonk Goes Open Weight: A Trillion-Parameter Bet on Sovereign Cyber Defense

Mistral has published the weights of Mistral Large 4, nicknamed “Le Chonk,” a trillion-parameter model that any organization can download and run on its own infrastructure. The French lab positions the release squarely around defensive cybersecurity and agentic coding, pricing its API at $0.70 per million input tokens and $2.10 per million output tokens. The weights landed on Hugging Face on October 29, three weeks after a public preview that Mistral says exceeded its expectations.

• Scale: Roughly 1 trillion total parameters with about 49 billion active per token, a 1M-token context window and native multimodality.
• Licensing: A modified MIT license that permits commercial use but requires a separate deal for anyone reselling the model as a service.
• Positioning: Mistral calls it the best open-weight model built in the West, while conceding that top Chinese open models still score higher on general capability.

Architecture: sparse experts, sparse attention

Le Chonk is a granular mixture-of-experts design. Its text decoder stacks 72 layers with a hidden dimension of 6,144 and 64 attention heads. The first layer uses a dense feed-forward network; the remaining 71 rely on MoE blocks where a softmax router selects four of 128 experts per token, alongside one shared expert that always runs. That yields five active experts per token and keeps inference costs far below what a dense model of the same size would demand.

Attention combines Multi-head Latent Attention with a DeepSeek-style sparse attention indexer that selects up to 2,048 key positions per query. A 3:1 RoPE-to-NoPE pattern alternates across decoder layers, and KV state is shared between them. Vision comes from a 1.6-billion-parameter Pixtral encoder using 14×14 patches. Mistral trained the model from scratch on roughly 3,800 Nvidia Grace Blackwell GPUs in its own European data centers, then spent three additional weeks of reinforcement learning after the preview.

Open weights do not mean lightweight. The files occupy close to a terabyte, and the model card suggests eight GPUs. Hugging Face’s guidance points to eight Nvidia H200 accelerators, or heavily compressed builds running on a single Mac with 512 GB of memory. Official FP8 and NVFP4 variants ship alongside community GGUF quantizations.

Cyber defense: strong refusals, second-tier capability

Mistral’s cybersecurity story is nuanced. On the Artificial Analysis Cyber Index, which tests whether a model can find, reproduce and patch vulnerabilities, Le Chonk scores 51%. That trails DeepSeek-V4.1-Flash at 57% and MiMo-V2.6-Pro at 56%, though it beats GLM-5.3 Flash, Kimi K3 and GLM-5.3. On the broader Artificial Analysis intelligence index it sits at 26 points, and the model card itself acknowledges that the strongest Chinese open models score higher.

Where the model does lead is safety. Mistral reports an average refusal rate of 96.6% on malicious cyber requests, the highest it has measured among open models, and resistance to 97.1% of attacks on Lakera’s B3 benchmark. In other words, it may not know the most about offense, but it is the hardest to weaponize.

“Concretely, it means that, for instance, as a business company or as a government, you’re able to have access to frontier intelligence to defend yourselves against cyberattacks. Access that can’t be cut off by definition because the model is open-weighted,” said Pierre Stock, Mistral’s Vice President of Science.

Mistral also acknowledges risk: testing showed the model declined to assist with blind attacks on real targets, but the company sees potential for misuse and says it is tightening safeguards. The specialized cybersecurity variant is available only in preview to vetted enterprise partners such as Orange Cyberdefense, limited to owned or authorized systems. ENISA, the EU cybersecurity agency, is trialing the model on vulnerability work, code audits and log analysis; its communications lead, Laura Heuvinck, argues open models let organizations retain control over the technology their resilience depends on. Hugues Foulon, CEO of Orange Cyberdefense, says defense must operate at a different scale against faster, increasingly automated threats.

Coding gains and the sovereignty argument

Beyond security, Mistral claims the best open-weight agentic coding model. On Cognition’s private FrontierCode 2.0 benchmark it scores 19.7%, and on Terminal-Bench 4.0 it reaches 41%, up 13 points from preview while cutting output tokens by 30%. On the Artificial Analysis Coding Agent Index it lands at 48%, effectively tied with GLM-5.3 while using 47% fewer tokens per task. Three reasoning modes — low, medium and high — let developers trade cost against precision.

The license is a modified MIT that allows commercial use, with one carve-out: offering the model to third parties through your own API requires a separate commercial agreement. Internal enterprise deployment remains free. Microsoft is the first cloud provider to offer it, and the model is also available through Mistral’s platform, OpenRouter, Hugging Face and Nvidia.

The release is the first milestone in a roadmap funded by Mistral’s €3 billion raise, and it arrives amid a wave of AI-assisted vulnerability discoveries. Le Chonk is genuinely good news for European sovereignty and for any organization unwilling to depend on a US API, and its pricing is aggressive. But framing it as a model that “beats China” would be misleading: by Mistral’s own cybersecurity chart, two Chinese rivals remain ahead on capability. What Mistral has built is a highly capable model that refuses better than its peers — arguably more valuable for defenders, provided the claim is told as it is.