Mistral AI launched Mistral Large 4 on October 6, 2026, a natively multimodal model with 1 trillion total parameters and 49B active, about 4.9% of the model per token. The announcement introduces it as "Mistral Large 4, aka Le Chonk", with an orange pixel-art Mistral logo hanging in a museum, cats in the paintings. It is available to all through the API today, with open weights promised for the end of October. VentureBeat, TNW, CNBC and Euronews all covered the launch the same day; CNBC's headline says Mistral claims it rivals the best open systems from China, and TNW frames it as a challenge to China's lead in open AI models.
What Mistral says it does best
Mistral calls Large 4 "the best open weights model from US or Europe on aggregated benchmarks". It describes the model as state of the art on critical workloads, naming cyber defense, manufacturing and finance.
The sharpest claim is visual grounding, where Mistral says Large 4 surpasses closed frontier models. It reportedly scores 73% on DIOR-RSVG, a remote-sensing benchmark, against 68% for GPT-6 Astra, a five-point gap. The model takes text and image input and outputs text.
How it ships
The API is open to everyone today through Mistral's API and Mistral Studio. For cyber, Mistral says it is working privately with cybersecurity partners. Press reports describe a separate private API with fewer restrictions and more cyber features for vetted cybersecurity firms and defense or state partners, with the French Ministry of the Armed Forces reportedly named as one.
The open weights are due on October 27, after about three weeks of testing with developers, cybersecurity leaders and government authorities. They will come in FP8 and FP4 formats; AlphaSignal's headline says the release "packs 1 trillion parameters into 4 GPUs". At that size, the memory problem described in Open Weights You Can't Actually Run is exactly what the low-precision formats are meant to ease.
Mistral stresses that the model was "forged in Europe end-to-end" and is deployable from Europe on its own Mistral Cloud infrastructure, the regional-inference business covered in Mistral Stops Selling Compute and Starts Selling Trust. Press reports say it was trained from scratch over roughly two months on about 4,000 Nvidia Grace Blackwell GPUs in Mistral's own European data centers.
Context
Large 4 follows Mistral Large 3, a 675B-parameter mixture-of-experts model released with open weights under Apache 2.0 in December 2025. Mistral Small 4, also Apache 2.0, arrived in March 2026. Total parameters have grown by roughly 1.5 times from Large 3, while active parameters sit at 49B.
The launch comes a month after Mistral closed its Series D of about 3 billion euros at a valuation above 21 billion euros, led by Samsung, which the company framed in its funding announcement as the basis for sovereign, open-weight AI at the frontier. It is the second European open-weights release this week, after Aleph Alpha's Kolibri, a 78B-parameter model with 3.46B active parameters.
The cyber positioning echoes other recent launches. Google gated Gemini 4 Argon to trusted testers and cyber defenders last week, and Anthropic's analysis of GLM-5.3 showed advanced cyber capabilities already spreading into open-weight models. Mistral's answer is to split the release: broad API access now, a less restricted private tier for vetted partners, and open weights only after a testing window with developers, security leaders and governments.