Mistral AI Unveils Mistral Large 4 with 1T Parameters
Mistral AI launched Mistral Large 4, a 1-trillion parameter multimodal MoE model. Now on API, open weights arrive late October.
Nicknamed “Le Chonk”, the multimodal MoE system runs first via commercial API, with downloadable weights promised for late October.
On October 6, 2026, French AI lab Mistral AI unveiled Mistral Large 4 (ML4), internally codenamed “Le Chonk”. The release marks the lab’s first push into trillion-parameter territory, integrating text and vision into a single system. Commercial access opened immediately through the Mistral Studio API, while downloadable open weights remain scheduled for release on October 27.
For developers and enterprises watching European AI infrastructure, the launch tests whether an open-weight model trained independently can match top proprietary systems from US cloud giants.
Architecture of Mistral Large 4: Sparse MoE and Vision
Rather than activating its full weight count on every run, Mistral Large 4 relies on a Mixture-of-Experts (MoE) design. In practice, MoE works like a team of specialized technicians where a central router directs incoming tasks only to relevant sub-networks. While the entire network houses 1 trillion parameters, each query routes through roughly 49 billion to 52 billion active parameters. This keeps compute costs and response latency manageable while preserving overall capacity.
Vision processing runs through an integrated 1.6-billion parameter vision encoder. The model comes with native pre-training across more than 160 languages, covering every official language of the European Union.
Mistral states that training ran from scratch inside European data centers on an in-house cluster of roughly 3,800 to 4,000 NVIDIA Grace Blackwell GPUs, bypassing direct reliance on US cloud providers. The technical push follows a €3 billion Series D funding round in September 2026 at a €21 billion valuation, led by Samsung with earlier backing from ASML.
API Pricing and the October 27 Weight Release
Immediate access is restricted to Mistral Studio’s paid API. The published rate sits at $1.36 per million input tokens and $4.18 per million output tokens. For context, a token is a word fragment roughly equal to four characters of text.
Teams waiting to self-host the model must wait until October 27. Mistral says the three-week gap allows for red-teaming and cybersecurity safety reviews with government agencies and security partners.
Two operational points remain unconfirmed: whether the public release will use a permissive Apache 2.0 license or restrictive commercial terms, and whether the downloadable public weights will carry the exact same capabilities and moderation boundaries tested by state agencies. Until Mistral publishes the license agreement, self-hosting terms remain a RUMOR.
Benchmark Claims and Pending Independent Audits
In internal benchmarks published by Mistral, ML4 posts scores outpacing proprietary closed models in visual grounding, scientific computing on SciCode-Verified, finance, and cybersecurity evaluations. The company positions it as the strongest open-weight AI built in the West.
Those performance figures currently represent self-reported claims from Mistral’s internal documentation. Independent verification and reproducible third-party evaluations remain unconfirmed until public weights ship on October 27.