Mistral Large 4 'Le Chonk' is now open weight: what it is and who can use it
Quick answer
Mistral Large 4, nicknamed “Le Chonk,” is Mistral AI’s biggest model: about 1 trillion parameters, multimodal (text and images in, text out), with a 1-million-token context. On October 29, 2026, Mistral released an improved version with downloadable weights, after a public preview earlier this month. It’s free to use, modify and sell, except businesses running a model-as-a-service need a separate license (Mistral).
The facts
- Released: Oct 29, 2026 (The Verge).
- Size: 1.05 trillion total parameters, 52 billion active per token (a Mixture-of-Experts design), plus a 1.6B vision encoder (Mistral docs). Hugging Face lists 49B active, which excludes embeddings and output layers (Hugging Face).
- Context window: 1 million tokens (Hugging Face).
- Where to get it: Mistral Studio, the Mistral API, OpenRouter, or the weights on Hugging Face for vLLM/SGLang; Microsoft Foundry is adding it (Mistral; Microsoft).
- API price: $0.70 per million input tokens, $2.10 per million output tokens (Mistral). Info only.
- Reasoning modes: low, medium and high (default).
How good is it? (Mistral-reported unless noted, Oct 29, 2026)
- Artificial Analysis Coding Agent Index: 48%, effectively tied with GLM-5.3 while using 47% fewer tokens per task.
- Terminal-Bench 4.0: 41%, up 13 points from the preview.
- Artificial Analysis Cyber Index: 51%.
- Malicious cyber requests: 96.6% average refusal rate, which Mistral calls the highest it measures among open models.
- Independent: Artificial Analysis scores it 26 on its Intelligence Index v5 at the high reasoning setting (up from 25 for the preview), the highest for an open-weight model built outside China (Artificial Analysis; VentureBeat). Chinese open models score higher overall; GLM-5.3 and DeepSeek V4.1 Flash lead open models at 32 (Artificial Analysis).
Why open weights matter
- You can download it and run it on your own hardware, so your data never leaves your servers. Mistral pitches this as “sovereign AI” for governments and regulated industries that don’t want their data routed to other countries.
- The trade-off: a 1-trillion-parameter model needs serious data-center GPUs. This isn’t a laptop model.
- Mistral says it saw “potential for uplift to conduct real attacks” in live traffic and is tightening guardrails. A specialized cyber variant goes only to vetted partners (Mistral). Compare Google’s gated approach with Gemini 4 Carbon.
The license, plainly
- The Mistral Ecosystem Open-Weight license, a modified MIT license: free to use, modify, fine-tune and commercialize.
- One condition: businesses that operate a Model-as-a-Service offering must get a separate commercial license from Mistral (Mistral). So “MIT-licensed” alone is not accurate.
FAQ
- Can I use Le Chonk for free? Yes for self-hosting under the license; the API is paid.
- Can I run it at home? Realistically no; it’s far too large for consumer hardware.
- Is it better than GPT or Claude? Mistral’s own comparisons are mostly against open models; in Surge’s blind coding test it ranked behind only Claude Opus 5.5 among models tested (Mistral-reported).
Sources
Primary sources
- Mistral: Mistral Large 4 (Le Chonk) is now open weight
- Mistral docs: model page
- Hugging Face blog
- Microsoft Foundry blog