Mistral Large 4 goes open weight with 1 trillion parameters
Anyone can now download and self-host Le Chonk, but companies selling it as a hosted service need a separate license from Mistral.
French AI company Mistral released the weights of Mistral Large 4 on Thursday (Oct. 29), making its 1-trillion-parameter flagship model, nicknamed "Le Chonk," free to download and run on companies' own hardware.
The model is a mixture-of-experts (MoE) design, meaning only a slice of the network runs for each token: 49 billion of its roughly 1 trillion parameters, or 52 billion counting embedding and output layers, according to the Hugging Face model card. It has a 1M-token context window and accepts text and images.
The weights come under what Mistral calls the Mistral Ecosystem Open-Weight license, "a modified MIT license." Mistral said in its announcement that the model is "free to use, modify, fine-tune, and commercialize," with one condition: businesses that run a model-as-a-service offering need a separate commercial license from Mistral.
What isn't known yet is how the model holds up outside Mistral's own tests. The benchmark scores in the launch post are reported by Mistral, and the company said a full technical report is "coming soon."
Built for coding and cyber defense
Mistral first opened Large 4 as an API public preview on Oct. 6 and promised the weights by the end of the month. The company said the released version had "three weeks of additional reinforcement learning" since then, with the biggest gains in coding and defensive cybersecurity.
On the Artificial Analysis Coding Agent Index, Mistral says Large 4 scores 48 percent, "effectively tied" with Z.ai's GLM-5.3, the leading open-weight model, while using 47 percent fewer tokens per task. It also reports 41 percent on Terminal-Bench 4.0, up 13 points from the preview.
In blind human-preference tests run by Surge, Mistral says graders preferred Large 4's code over GLM-5.3, Moonshot's Kimi K3 and DeepSeek V4.1 Flash, with only Anthropic's Claude Opus 5.5 ranking ahead of it among the models tested.
On security, Mistral reports 51 percent on the Artificial Analysis Cyber Index, which tests whether a model can find, reproduce and patch vulnerabilities. It says the model refused 96.6 percent of malicious cyber requests across three safety benchmarks, the highest rate it measured among open models.
Pierre Stock, Mistral's vice president of science, framed the release as a security argument. "Access that can't be cut off by definition because the model is open-weighted," Stock told TechCrunch, describing what governments and companies gain.
Mistral also flagged a risk in its own post. After monitoring cyber-related use during the preview, it said "we see a potential for uplift to conduct real attacks," and that it is "modifying our guardrails accordingly." A specialized variant for penetration testing and red-teaming is going only to vetted enterprise partners such as Orange Cyberdefense.
On broader tests, the picture is weaker. The model card puts Large 4 at 26 on the Artificial Analysis Intelligence Index v5, which Mistral calls the highest among Western open-weight models while noting that the leading Chinese open models score higher.
What it takes to run
Le Chonk is not a laptop model. Mistral ships FP8 and NVFP4 checkpoints, and the vLLM project's day-0 recipes run the FP8 version on 8 GPUs and the NVFP4 version on 4 Nvidia Blackwell GPUs. On AMD hardware, vLLM says a requantized checkpoint shrinks from about 1,061 GB to 578 GB and fits on four MI350 Series GPUs.
Support landed on day one in vLLM, SGLang, Hugging Face Transformers and llama.cpp, with community GGUF files from Unsloth and the llama.cpp team, according to Hugging Face. The checkpoint includes a built-in speculative decoding head, which Mistral says raised throughput by up to 1.9 times in its tests.
For teams that don't want to host it, the API costs $0.70 per million input tokens and $2.10 per million output tokens, with cached input at $0.07, per Mistral's pricing page. The model is also on OpenRouter, Nvidia-hosted endpoints and Microsoft Foundry, which Microsoft says makes it the first cloud provider to offer the model.
The release lands in a crowded week for security-focused models. Google said Wednesday that Gemini 4 Carbon will go to vetted cyber defenders first, and Moonshot has promised the Kimi K3.1 Pro weights by Nov. 4. Mistral said Large 4 will serve as the foundation for a new generation of specialized Mistral models.
