Mistral Ships 1-Trillion-Parameter Open-Weight Model Built on Its Own European Cluster

Mistral AI opened a public preview of Mistral Large 4 on October 6, putting a 1-trillion-parameter model into developer hands and promising open weights by the end of October. The move is a meaningful infrastructure story as much as a model story, and the infrastructure part is what's worth paying attention to.

The architecture is mixture-of-experts: 1 trillion total parameters, but only 49 billion active on any given forward pass. That's the design decision that makes a trillion-parameter model usable at inference without proportional compute cost. As reported by Artificial Intelligence News, the preview API is live now in Mistral Studio, with weights and full architecture details scheduled for release at month's end.

What separates this launch from most large-model announcements is where the compute came from. According to Artificial Intelligence News's coverage of Mistral's announcement, the company trained the model from scratch on its own cluster of 3,800 NVIDIA Grace Blackwell GPUs in its own European data centers. That's a deliberate sovereignty bet. Mistral isn't renting time from AWS or Azure; it's running its own iron in Europe. The company even highlighted in its announcement that the model is "forged in Europe end-to-end," language that tracks closely with how European regulators and government customers think about supply-chain trust.

The predecessor, Mistral Large 3, launched in December 2025 with 675 billion total parameters and 41 billion active, according to OfficeChai. The jump to a trillion parameters and full multimodality is a real generational step for the lab, not a point release.

The model supports over 160 languages, including every official EU language, according to Artificial Intelligence News. It's natively multimodal, and Mistral is positioning cybersecurity as a primary use case. Per Artificial Intelligence News, the company says Large 4 solves 93 percent of Cybench's 40 security-competition exercises, and Mistral has been running private access with cybersecurity partners and state authorities before the general release. CEO Arthur Mensch told a conference in Abu Dhabi that the model outperforms Chinese models on cybersecurity benchmarks, according to CNBC.

On independent evaluations, the picture is more modest. Simon Willison, reviewing the model via Artificial Analysis benchmarks in his link blog, noted it scored 38 on Artificial Analysis, placing it just behind DeepSeek 4.1 Flash, which is a 552-billion-parameter model. His read: the model is "maybe about 6 months behind the frontier." That's a useful data point. Mistral's internal benchmarks and an independent evaluator's quick tests don't fully align, which is normal at preview stage, but worth flagging before enterprises plan workloads around the cybersecurity claims.

Pricing lands at $1.36 per million input tokens and $4.18 per million output tokens, per OfficeChai's coverage of Mistral's announcement. That's positioned above commodity open-weight inference but well below frontier closed-model rates.

The license question is still open. Mistral's recent releases have been mixed: Large 3 and Small 4 shipped under Apache 2.0, while Mistral Medium 3.5 arrived under a modified MIT license, according to OfficeChai. The company hasn't named the terms for Large 4's open release yet. For enterprises that want to self-host, fine-tune, or audit the model, that matters more than the parameter count. A trillion-parameter model under a restrictive commercial license isn't really an open-weight model in the ways that drive enterprise adoption.

The weights drop date Axios reported is October 27. That's when the architecture writeup and post-training methodology are also supposed to arrive. Until then, developers are working against a black box with an API in front of it, which limits how much of Mistral's cybersecurity story can actually be verified.

Sources cited:
- Artificial Intelligence News (https://www.artificialintelligence-news.com/news/mistral-ai-launches-large-4-preview-ahead-open-weight-release/)
- OfficeChai (https://officechai.com/ai/mistral-large-4-le-chonk/)
- Simon Willison's Weblog (https://simonwillison.net/2026/Oct/6/le-chonk/)
- CNBC (https://www.cnbc.com/2026/10/06/mistral-ai-model-le-chonk.html)