The model has a nickname inside the company: “Le Chonk.” The name is a joke, but the ambition behind it is not. Mistral AI, the French startup that has spent three years arguing that the most capable AI should not be locked inside the servers of a couple of American firms, on Monday unveiled Mistral Large 4, a model with a trillion parameters that it says matches the best closed systems on the market and will be handed over, weights included, by the end of the month.
Large 4 is built on a mixture-of-experts design that spreads those trillion parameters across many specialized subnetworks and activates only about 49 billion of them for any given query. The result, Mistral says, is a model that is cheaper to run than a dense system of similar power while matching frontier performance on the tasks businesses actually buy.
The claims are specific. On the Artificial Analysis intelligence index, the company says Large 4 ties GPT-6 Luna, OpenAI’s latest flagship, and leads in subtests covering law and mathematics. On a cyber index that measures how well a model finds and patches security flaws in real software, it scores 82% on a test that asks a model to reproduce a known vulnerability and then fix it, the highest mark of any model Mistral cites. The figure is aimed squarely at the people who approve security budgets.
The training story reinforces the pitch. Mistral says Large 4 was trained from scratch on roughly 3,800 Nvidia Grace Blackwell GPUs in data centers the company runs itself in Europe. The detail doubles as a sales argument: enterprises and governments that do not want their AI running on American infrastructure can buy a model that never left the continent, handled end to end under European law.
The release is staged rather than sudden. Large 4 is not yet open. It launched in public preview through Mistral’s API, priced at $1.36 per million input tokens and $4.18 per million output tokens. The weights, the files that let a customer download and run the model on their own hardware, arrive at the end of October. Mistral has used this gap before: it opens the preview, courts cybersecurity firms and government agencies with a more permissive private version, and releases the open weights once the enterprise pipeline is primed.
That two-tier approach has become the company’s signature. The public API carries standard safeguards. A separate, more lightly moderated version goes to vetted partners, mostly cybersecurity companies and state authorities, with broader capabilities. It is the same playbook Mistral has run for earlier models, and it positions the firm as the open-weight alternative that still knows how to sell to institutions.
The market Large 4 targets has shifted over the past year. Buyers that once treated open-weight models as research toys now want them for production, partly to escape dependence on OpenAI and Anthropic and partly because regulators in Europe have begun asking where data and compute actually live. Mistral argues that a trillion-parameter model, self-hosted and open, answers both concerns at once.
The company says Large 4 is natively multimodal, taking text and images as input, and answers in more than 160 languages. That breadth is aimed at the enterprise contracts that keep a business alive, not the leaderboard churn that fills a launch day. Mistral’s own filings position the model for cybersecurity, finance, manufacturing, and law, the corners of the market where auditable, self-hosted software carries the most weight.
None of this means the closed labs are losing. OpenAI and Anthropic still lead on the most popular coding benchmarks, and Mistral’s own numbers show Large 4 trailing frontier systems there. The question Mistral is forcing is different: whether, for the slice of enterprise work that involves security, legal analysis, and finance, an open model that can be audited and run in-house is good enough to win the contract.
Mistral’s founders built the company on a bet that Europe needed its own answer to the American labs. Arthur Mensch, the chief executive, has argued for years that models running entirely on European infrastructure and under European law are not a luxury but a requirement for governments and regulated industries. Large 4 is the fullest expression of that bet so far: a trillion-parameter model trained in the company’s own data centers, released to the public rather than kept behind an API. Whether the market wants a European open model at scale, and whether a startup can afford to keep training them, are questions the next quarter will begin to answer.
Mistral is betting the answer is increasingly yes. And the fact that the strongest open model built outside China now carries a joke nickname is itself a kind of argument. The company wants to look like the pragmatic underdog, not the lab chasing a science-fiction prize. The trillion parameters are the engineering; the posture is the marketing.
