Mistral Large 4 Trails Chinese Open Models, Bets on Weights

A trillion parameters, with a qualifier

Mistral AI opened a public preview of Mistral Large 4 on October 6. The model's in-house nickname is "le Chonk," and the French lab calls it its largest and most capable model so far 23. It is a natively multimodal mixture-of-experts system with roughly a trillion total parameters and about 49 billion active per token 24. Right now it is available only through Mistral's API, under the model ID mistral-large-4 3.

The company's own claim deserves a close look. Mistral describes Large 4 as the most capable open model outside of China 5. That phrasing admits that the strongest open-weight systems still come from Chinese developers. CNBC described access to those Chinese open models as a key flashpoint in the US-China race for AI supremacy, and noted that they are increasingly being adopted worldwide 5. Mistral is not claiming the open-model crown. It is claiming the best seat available to a Western, and specifically European, lab.

The spec sheet doesn't fully agree with itself

Several of the basic figures differ depending on where you look:

• Parameter count. The model card lists 1.05 trillion total parameters, while the announcement rounds this to "1 trillion." 1
• Active parameters. The announcement says 49 billion are active per token, but the model card says 52 billion. 1
• Context window. Mistral's card lists one million tokens, while Artificial Analysis records 524k for the preview API. 1
• Training compute. One account cites 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European datacenters 1. Another gives a range of roughly 3,800 to 4,000 4.

None of these gaps is dramatic. Still, the context-window difference is the one buyers should notice. A model advertised at a million tokens but served at about half that in preview is a practical limit, not a footnote. The model also includes a 1.6-billion-parameter vision encoder and accepts text and images while returning text 1.

Pricing and the "open" question

The listed API rates are $1.36 per million input tokens, $0.14 for cached input, and $4.18 per million output tokens 1. During the preview, Mistral is charging roughly half of that: $0.68 for input, $0.07 for cached input, and $2.09 for output 12.

The weights are the bigger story, and they are not out yet. Mistral says they will "drop" by the end of October 2. One outlet pins the date to October 27 4. The licence terms are still undisclosed 13. Mistral's model page labels the licence "Open" without explaining what that means 2. This matters because the label "open weights" covers everything from permissive licences to restrictive custom terms. Until the terms are published, enterprises cannot judge whether le Chonk actually competes with the Chinese models it is being compared to.

A cybersecurity-first rollout

The release is staged in a notable way. Mistral says Large 4 is particularly strong at cyber, coding, manufacturing, finance, and multimodal tasks 5. The early preview went to developers, to "cybersecurity leaders," and to state authorities before a wider release later in the month 5. Basic Tutorials reports that the open weights will follow only after additional security testing is finished 3. It also points out that organizations wanting to run the model on-premises or in a private cloud for security work will have to wait, because self-hosting is impossible until then 3. Explainx.ai's coverage refers to the model's DeepSWE and Cybench results, the latter being a cybersecurity-oriented benchmark 4.

The likely logic is that a frontier-scale model with strong offensive and defensive cyber capabilities is risky to release openly. Giving defenders and governments a head start, and testing before the weights ship, is a reasonable hedge. It also signals that Mistral expects scrutiny of what the model can do once anyone can download it.

The verdict: chosen for sovereignty, not supremacy

Beam.ai's headline puts it well: this is "a model you choose for the weights" 1. Every source agrees on the trillion-parameter scale and the promise of openness 12345. Mistral's own framing concedes that Chinese open models remain ahead 5. So the case for Large 4 is less about leading benchmarks and more about where it comes from. It was trained in European datacenters 1, built by a European company, and is intended to be self-hostable.

For governments and security teams that are wary of building on Chinese foundations, that may be enough. Whether it is depends on details that have not been released: the licence, the actual context limit, and whether the weights arrive on schedule. Until those are confirmed, le Chonk is a promising preview, not a finished open model.