Mistral unveils Large 4, a 1.05-trillion-parameter open-weight model, with the weights still to come

The news

Mistral announced Mistral Large 4 on 6 October 2026, a model it has nicknamed “Le Chonk”. It is a natively multimodal mixture-of-experts with 1.05 trillion total parameters and about 52 billion active per token (the released artifact is named Mistral-Large-4.0-1T05-A52B; some early coverage cited around 49 billion active), a 1.6-billion-parameter vision encoder, and a context window of one million tokens. Mistral says it was trained from scratch on roughly 3,800 NVIDIA Grace Blackwell GPUs in the company’s own datacentres in Europe, across more than 160 languages.

The release is staged. The preview is available through Mistral’s API now, priced at 1.36 US dollars per million input tokens and 4.18 dollars per million output tokens. The open weights are not yet downloadable: Mistral says it will release them by the end of October 2026.

Why it matters, and for whom

For the open ecosystem this is a notable release: a European lab shipping a frontier-scale open-weight model, trained on European infrastructure, with a strong multilingual emphasis. For most people running AI locally, though, it is worth being clear about scale. At 1.05 trillion parameters, even a 4-bit quantisation would be very roughly 500GB (an approximate figure, and multi-GPU server territory), so this is not a model you will run on a laptop, a single card, or even a maxed-out unified-memory workstation. Its relevance to local AI is as an option for organisations with serious server hardware, and as a sign of where open models are heading, rather than as a daily local driver.

The licence is the open question

Mistral describes Large 4 as “open-weight”, and the weights will be downloadable. That is not the same as a licence that permits commercial use, and as of this announcement Mistral has not published the licence the weights will carry. This matters, because Mistral’s recent large models have shipped under the non-commercial Mistral Research License rather than a permissive licence like Apache 2.0. Until the licence is published, treat the commercial terms as unknown, and read them at the primary source before building on the model. We have seen before that an “open” headline can hide real restrictions, including ones tied to where you operate; our piece on open models that exclude the UK covers why the fine print is worth reading.

The benchmarks are Mistral’s own, for now

Mistral reports strong figures on enterprise and security workloads, including 93 per cent on Cybench’s security exercises and 61.7 per cent on DeepSWE v1.1, alongside 28.3 per cent on Terminal-Bench 4.0 (Mistral’s own figures, 6 October 2026). Treat these as claims from the releasing organisation until independent evaluations appear. One early independent read already exists: the benchmarking firm Artificial Analysis placed Large 4 on its intelligence index between DeepSeek V4.1 Flash and a US frontier model, which puts it roughly level with the strongest Chinese open models and behind the leading US closed systems. That is a credible position for an open model, stated without the vendor’s framing.

The history

Large 4 lands in a run of very large open-weight mixture-of-experts models through 2026: Zhipu’s GLM-5.3, DeepSeek V4.1 Flash, and Tencent’s Hunyuan Hy4. The common thread is that frontier-scale open weights increasingly demand datacentre hardware to run, even as smaller models in the 20 to 35 billion range remain the practical choice for local use. Mistral’s distinctive angle here is European training and a multilingual focus, which is part of a broader sovereignty argument for open weights.

We will add Mistral Large 4 to the model catalogue once the weights are released and the licence is confirmed at the primary source, in line with our practice of cataloguing what you can actually run and verify.