Mistral Large 4: 1T Open-Weight Model Preview, Oct 2026

Europe answers back on open-weight AI. On October 6, 2026, Mistral launched a public preview of Mistral Large 4 (ML4, nicknamed “le Chonk”): a natively multimodal model with 1 trillion parameters, of which 49 billion are active, and Mistral’s largest model to date.

What Mistral Large 4 is

• A text-and-image model aimed at coding, agentic workflows and multimodal understanding
• Trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral’s own European datacenters, which also serve the preview (The Next Web reports about 4,000 GPUs over two months)
• Training data spanning 160+ languages, including every official EU language
• The first model funded by Mistral’s €3 billion Series D

Benchmarks (vendor-reported, preliminary)

Mistral lists 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA and 28.3% on Terminal-Bench 4, plus 59.9% on AutomationBench. On security, it reports 93% of Cybench challenges solved and 82% on a reproduce-and-patch test of the Artificial Analysis Cyber Index. The reinforcement-learning run is still in progress, so these numbers are expected to move before the weights ship.

Sovereignty and cybersecurity

Mistral pitches ML4 as a model organisations can run on private cloud or on-premise, with a European deployment it operates under EU law. Before releasing the weights, it is red-teaming the model with cybersecurity leaders, vetted partners and state authorities.

Availability

The preview API is live on Mistral Studio. Mistral says the weights, plus architecture details, more benchmarks and its post-training method, will follow by the end of the month; The Next Web gives October 27.

Rocks take

A 1T-parameter model is not something you run next to Home Assistant on a mini-PC, but open weights from a European lab matter for anyone building sovereign, self-hosted AI stacks. Until independent tests land, treat every score above as Mistral’s own claim.

FAQ

Can I try it now? Yes, via the preview API on Mistral Studio.

Is it multimodal? Yes: text and images.