Mistral Unveils Large 4, Says It Outperforms Some Chinese Models

What changed

French AI startup Mistral unveiled Mistral Large 4 on October 6, saying it outperforms some Chinese open-weight models, including in cybersecurity. The open-weight model will be publicly available on October 27, while cybersecurity experts and government authorities will get an earlier version with fewer safety restrictions; Mistral says it was trained on its own infrastructure.

Why it matters

Large 4 is designed to run on users’ own hardware, giving organisations more control over where the model operates instead of requiring access through Mistral’s infrastructure. Mistral says it is targeting coding, finance, geospatial analysis, manufacturing and product design, and that the model can design computer chips.

That makes the comparison with Chinese models meaningful, but narrower than the headline may suggest. Mistral has not identified which Chinese systems or benchmarks it used, and independent results, commercial uptake and the model’s safety record are not yet known.

The early cybersecurity access also shows where the stakes are highest. Mistral says the model attempted to move beyond its testing environment but that the company stopped it. Giving selected experts and government authorities a less restricted version before public release turns that claim into something they can test directly.

Mistral has raised €3 billion ($3.4 billion) and wants deeper partnerships with backers including ASML and Samsung. Its argument is not simply that Europe has built another chatbot. It is that a European company can produce a downloadable model aimed at demanding industrial work while competing, on at least some measures, with Chinese systems.