On October 6, 2026, Mistral AI announced Mistral Large 4, affectionately nicknamed "Le Chonk" — and the name is earned. The model packs roughly 1 trillion total parameters, with 49 billion active at inference time through a mixture-of-experts architecture, plus a 1.6-billion-parameter vision encoder that makes it natively multimodal from day one.
The API is live now through Mistral Studio. Open weights are planned for before the end of October — which would make Le Chonk the largest openly released AI model by a substantial margin.
What's inside
Mistral Large 4 was trained from scratch on approximately 3,800 Nvidia Grace Blackwell GPUs located in Mistral's own European data centers. No US cloud dependency. It supports over 160 languages, including every official EU language, and handles documents, charts, technical drawings, and satellite imagery alongside plain text. The context window is 1 million tokens.
On cybersecurity benchmarks, Mistral reports 93% on Cybench and 82% on CyberGym-E2E. Coding and agentic workflows are listed as key strengths. Independent third-party benchmarks are still incoming, but early community reports on code generation have been positive.
Pricing via API sits at $1.36 per million input tokens and $4.18 per million output tokens — competitive with GPT-6 and Claude tiers, and effectively free once the weights drop for anyone who can run them locally.
The open-weights angle matters
In 2026, the AI landscape splits cleanly between closed models and open ones. Mistral's previous releases — Mixtral 8x22B, Mistral Large 2 — were genuine open-weight models that researchers, startups, and privacy-conscious enterprises relied on as an alternative to API-only giants.
Le Chonk, if it releases as promised, would be the first trillion-parameter-class model available for local deployment. That's a qualitative shift: capable enough to compete with frontier models, but runnable on your own infrastructure with no data leaving your environment.
For academic research, healthcare, legal, and government applications where data sovereignty is non-negotiable, this is significant.
European AI sovereignty in practice
TechCrunch's coverage notes that third-party data currently places Le Chonk below the leading closed models, and full benchmark results are still pending. Mistral hopes it will be best-in-class among open-weight models — a more realistic bar than matching GPT-6, and a meaningful one if it holds.
Training in France, releasing weights openly, and supporting EU languages aren't just marketing. They position Mistral as a structural counterweight to US-centric AI infrastructure at a moment when the EU's AI Act is making data provenance and model governance genuinely consequential for enterprises operating in Europe.
The open weights aren't out yet as of this writing. When they arrive, expect the self-hosting and open-source ML communities to run hard on them. Le Chonk is the most anticipated open model release of 2026 — and the name alone will make sure nobody forgets it.