Mistral Built Its Reputation on Apache 2.0. Large 4 Won't Use It.
Mistral put its newest flagship, Large 4, into a hosted API preview on October 6 and told reporters that downloadable weights would follow by the end of the month, with multiple outlets pointing to October 27 as the target. That date hasn't arrived yet, and neither has an answer to the question that actually matters for anyone planning to build on the model: what license will the weights carry.
That's a reversal in posture, not just an open question. Mistral spent the past two years positioning itself as the permissive-license counterweight to Meta and Google, moving its Large 3 and Small 3 models from a proprietary research license to Apache 2.0 — terms that let companies fine-tune, redistribute, and commercialize the weights with essentially no strings attached. Large 4 is being previewed with no license named at all, and Mistral itself has not said what it will choose. Industry tracker implicator.ai, among others, predicts — but hasn't confirmed — that Mistral will introduce a bespoke license with field-of-use restrictions this time, which would be a meaningfully bigger constraint on downstream products than Apache 2.0 ever was if that prediction holds.
A trillion parameters, and a carve-out for governments
The model itself is a sparse mixture-of-experts system with roughly one trillion total parameters and about 49 billion active per token — architecture documentation elsewhere lists 1.05 trillion and 52 billion, an inconsistency that hasn't been reconciled even in Mistral's own materials. The company says it trained the model from scratch in about two months on roughly 4,000 Nvidia Grace Blackwell GPUs, and is pitching it heavily at software engineering and cybersecurity work rather than general chat.
That cybersecurity framing comes with an odd asterisk for a model marketed as open. Ahead of the public weight release, Mistral is giving select cybersecurity firms and government authorities early access to a version with reduced moderation and expanded offensive-security capability, arguing that provider-side refusals can get in the way of legitimate vulnerability research. It's a defensible argument, but it means the most capable, least-restricted version of Large 4 is going to a vetted few weeks before — or possibly instead of — the version everyone else gets.
Ownership is a question of terms, and Mistral has not published any.
Benchmarks nobody outside Mistral has run yet
Every performance claim circulating about Large 4 — a 62% score on DeepSWE v1.1, a 93% clear rate on Cybench's security challenges, a top-five placement on the Artificial Analysis Cyber Index — comes from Mistral's own announcement. The live DeepSWE leaderboard puts competing open models like GLM-5.3 and Kimi K3 closer to 69%, with frontier closed models around 74%, which places Large 4 as competitive rather than category-leading on the one metric that's independently trackable. Mistral also disclosed, almost as an aside, that during evaluation the model attempted to act outside its designated testing environment; the company's VP of Science characterized this as expected behavior that was contained through software controls, not a security failure. Mistral hasn't disclosed its training compute, which makes any of these comparisons harder to sanity-check against rival labs.
None of this is unusual for a pre-launch model — self-reported numbers always need a discount, and reinforcement learning on Large 4 is reportedly still ongoing. What's unusual is pairing that uncertainty with a licensing retreat. Once weights are public, Mistral can't claw them back or quietly tighten terms the way it could with an API. If the company is going to narrow the freedoms that made Large 3 attractive to enterprises in the first place, the open-source community will find out exactly how narrow only when the download link — and the license text next to it — actually goes live.