Mistral Large 4 is now open weights: a 1-trillion-parameter model you can download

Mistral released the weights of Mistral Large 4 on October 29. The model, which Mistral calls Le Chonk, first appeared as an API-only public preview on October 6. Mistral says the version released today has had about three more weeks of reinforcement learning since then, mostly aimed at coding and defensive cybersecurity.

What it is. A mixture-of-experts model with about 1 trillion total parameters and 49 billion active per token (52 billion counting embedding and output layers): 128 routed experts, 4 active per token, plus 1 shared expert. It takes text and images in and writes text out, has a 1M-token context window, and supports function calling, JSON output and strict JSON schemas. Mistral says it was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own data centers in Europe.

Three reasoning levels. You can set reasoning effort to low, medium or high (the default) per request, so the same model can do quick replies or long agent runs. Mistral says low replaces the old "none" mode.

Mistral's numbers. On Terminal-Bench 4.0, Mistral reports 41%, up 13 points from the preview, while using 30% fewer output tokens. On the Artificial Analysis Coding Agent Index it reports 48%, about tied with GLM-5.3 but using 47% fewer tokens per task on average. On Cognition's private FrontierCode 2.0 benchmark it reports 19.7%, the top score Mistral shows. For security work, Mistral reports 51% on the Artificial Analysis Cyber Index and a 96.6% average refusal rate on malicious cyber prompts. On the broader Artificial Analysis Intelligence Index v5 it scores 26; Mistral's own model card says the leading open models from China score higher there. MODEL has not run any of these benchmarks.

How to get it. Weights are on Hugging Face as mistralai/Mistral-Large-4-1T-A52B, with an NVFP4 quantized version for Blackwell GPUs. Mistral recommends vLLM and also documents SGLang, Transformers, llama.cpp, and fine-tuning with Axolotl, Unsloth and TRL. It's also on the Mistral API as mistral-large-4 at $0.70 per million input tokens, $0.07 per million cached input tokens and $2.10 per million output tokens, plus OpenRouter, Mistral's Vibe coding agent, NVIDIA-hosted endpoints and Microsoft Foundry. Mistral says its official demo Spaces on Hugging Face have a small shared free budget per visitor for the first week.

Who it's for: teams that want a large open model they can host themselves, especially for coding agents, security triage, or data that can't leave their own servers. If you just want to try it, the API or the demo Spaces are the easy way in.

The caveats. This is not a laptop model: the official vLLM and SGLang recipes spread it across 8 GPUs. The license is a modified MIT license that allows commercial use, but a business that runs a Model-as-a-Service offering needs a separate commercial license from Mistral. The benchmark results are Mistral's, and FrontierCode is a private benchmark. Mistral also says that, watching live traffic since the preview, it saw "a potential for uplift to conduct real attacks" and is changing its guardrails.

What MODEL did not test: we didn't download or run the weights, call the API, or check any benchmark or pricing claim ourselves.

Not sponsored. No affiliate links.