What the announcement commits Mistral to
Mistral's October 6 post launches a public preview of Mistral Large 4 through the API on Mistral Studio and says the weights drop at the end of the month, which the dateline makes October. The company calls the model the first milestone on a roadmap funded by its EUR 3 billion Series D, which it describes as the largest equity round ever raised by a European technology company. That is Mistral's account of its money and its calendar. It establishes neither that the weights exist for download today nor that the funding superlative or the performance claims have been checked by anyone else.
Source: Introducing Mistral Large 4 · 2026-10-06 · Primary source context · attributed account
Follow the compute
Mistral says Large 4 is a natively multimodal mixture-of-experts model with one trillion parameters, 52 billion of them active, trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own European datacenters, with the preview served on the same hardware. It says the capital is already being put to work scaling up compute capacity in its own European datacenters, that the reinforcement learning run behind the preview is still in flight, and that the model will be the base for a new generation of specialized Mistral models. Silver Owl view: capital, equipment, a training run and a product are four different commitments. Only the preview can be examined today. The rest is roadmap, which is what a funding announcement is for.
Source: Introducing Mistral Large 4 · 2026-10-06 · Primary source context · attributed account
Who ran which test
The benchmark section rewards a close read of pronouns. Mistral cites the Artificial Analysis Cyber Index, which it calls an independent evaluation, and reports an 82 percent score on that index's reproduce-and-patch test. It reports solving 93 percent of the 40 Cybench challenges without saying who ran them. It says it ran a blind human evaluation of coding quality with Surge AI, in which annotators scoring one to five put the Large 4 preview second of five at 3.74, behind Claude Opus 5 at 4.22. It says third-party evaluator Vals AI tested legal and financial tasks. And it reports resisting 93.3 percent of attacks on Lakera's public B3 benchmark, while the indirect prompt injection claim rests on Mistral's own benchmarks. Every number arrives through Mistral's post.
Source: Introducing Mistral Large 4 · 2026-10-06 · Primary source context · attributed account
A deadline you can check
Silver Owl view: the useful thing about a promised date is that it expires. Mistral says the weights will come with architecture details, additional benchmarks and its post-training methodology. When October ends, the question is whether those artifacts arrived and what they let a reader verify. An illustrative exercise: run the API preview on one defined task, write down what it did, and keep a second column for everything that stays pending until the weights and documents land. Treat the benchmark rankings as company claims throughout. The funding headline is a statement of ambition. Delivery needs its own paperwork.
Source: Introducing Mistral Large 4 · 2026-10-06 · Primary source context · attributed account