Mistral Large 4 FAQ

Mistral Large 4 (ML4) is a mixture-of-experts (MoE) frontier language model announced by Mistral AI on October 6, 2026. It has 1.05 trillion total parameters with 49 billion active per token, a 1.6 billion-parameter native vision encoder for multimodal input, a 1 million-token context window, and support for 160+ languages. It is a hybrid model combining instruction following, reasoning, and agentic capabilities. See our full overview in the what-is-it guide.

The formal name is Mistral Large 4, abbreviated ML4. "Le Chonk" is the official nickname that Mistral AI itself used in its own launch announcement — an informal reference to the model's trillion-parameter size. So the nickname is official, but the product name is Mistral Large 4.

It was announced on October 6, 2026, alongside a public preview release through Mistral's API. The API model id is mistral-large-4, documented in API docs v26.10, with a playground available on the docs/models page. See the changelog for the full release timeline.

No. The weights are not downloadable today — Mistral Large 4 is API-only until the weights ship. The open weights are expected at the end of October 2026. Check the download page for the current status.

The open weights are expected at the end of October 2026. Reuters reported October 27, 2026 as the expected date, and Mistral plans to publish the weights on Hugging Face. The rollout will be staged: developers, cybersecurity professionals, and government authorities get access first.

The weights are expected to be released under a custom Mistral license. The release is planned as open-weight, not OSI open-source, so the exact permissions will depend on the terms of that custom license once published.

Preview pricing is live now: $1.36 per 1 million input tokens and $4.18 per 1 million output tokens. See the pricing page for details and a cost calculator.

Mistral Large 4 supports a 1 million-token context window.

Mistral claims it is the best open-weight model from the US or Europe on aggregated benchmarks, and competitive with the strongest open models worldwide. CEO Arthur Mensch added that it is above Chinese models on certain aspects, including cyber. Independent testing by Artificial Analysis places the ML4 preview between DeepSeek V4.1 Flash and OpenAI's GPT-6 Luna. For head-to-head detail, see our DeepSeek V4 Pro comparison, our Qwen 3.8 Max comparison, and our page on how it differs from closed frontier models.

It was trained from scratch — with no distillation — over roughly two months on about 3,800 Nvidia Grace Blackwell GPUs in Mistral's own European datacenters. The effort reflects how much the team has grown: Mistral started with a science team of 3 people and now has around 300 researchers, led by CEO Arthur Mensch and co-founder and chief scientist Guillaume Lample.

More than 160 languages, including all official EU languages.

Not in the OSI open-source sense. Mistral Large 4 is planned as an open-weight release: the weights will be published (expected on Hugging Face) under a custom Mistral license, but that is not the same as OSI open-source licensing. And until the weights ship, the model is API-only.

Through Mistral's API, which is live now in public preview. Use the model id mistral-large-4 (API docs v26.10), or try the playground on the docs/models page. See the quickstart guide for setup steps.

Developers, cybersecurity professionals, and enterprises or governments that want a self-hosted sovereign EU model. Mistral pitches it together with sovereign infrastructure and zero-data-retention options.

Mistral reports 93% on Cybench and 82% on CyberGym-E2E. Two caveats: these are vendor-reported scores, not independent results, and closed frontier models score near zero on these benchmarks because they refuse the tasks. See the benchmarks page for full context.