Mistral Large 4 Preview: Pricier API, Open Weights Due Oct. 27

What Mistral launched

On October 6, French AI company Mistral opened a public preview of Mistral Large 4 (ML4), its new flagship model. Internally the model carries the codename "Le Chonk." 14 It is a sparse mixture-of-experts system with roughly one trillion total parameters and 49 billion active per token. 15 One detailed spec breakdown puts the total at 1.05 trillion and lists a 1.6-billion-parameter vision encoder. 4 The model takes multimodal input natively and supports more than 160 languages. 1 Its context window grows to one million tokens, up from 256,000 in Mistral Large 3. 34

For now, access is limited to an API. Downloadable weights are scheduled for October 27, after a staged testing period. 24 Some coverage gives the timing more loosely as "end of October" or "month-end," which fits the same window. 13 Until then, the model goes to developers, cybersecurity leaders, and state authorities. 5 Mistral said security experts and government bodies would get a version with fewer safety restrictions so they can probe what it can do before the public release. 2

The sovereignty and China framing

Mistral is promoting ML4 mainly as a geopolitical statement. The model was trained on 4,000 Nvidia Grace Blackwell GPUs over about two months, inside Mistral's own European data centers. 5 The company says its European deployment runs end-to-end without reliance on third-party digital service providers and complies with EU law. 1 By parameter count, it is described as the largest open-weight model ever released from Europe. 1

The comparison Mistral keeps returning to is China. Most of the strongest open-weight models come from Chinese labs, while leading US models are usually closed and reachable only through their developers' own infrastructure. 2 Mistral calls ML4 the strongest open-weight model built outside China "by a substantial margin." It also expects the model to rank among the top open-weight systems worldwide on aggregate benchmarks once the weights are out. 5

Speaking at a launch event in Abu Dhabi, CEO Arthur Mensch said the model beats Chinese models "on certain aspects, including cyber." He did not say which models or which benchmarks. 2 Pierre Stock, Mistral's VP of science, said ML4 is among the strongest open-weight systems available. 2

Where the claims get softer

The coverage does not fully agree on how capable the model is. Mistral highlights strength in cyber, coding, manufacturing, finance, and multimodal tasks, and the company also concedes that ML4 still trails the frontier in areas such as coding. 5 These positions can coexist: a model can do well on coding among open-weight peers and still fall short of the best closed systems. Readers should keep that distinction in mind. Mensch's comparison to Chinese models is also hard to evaluate without named benchmarks. 2 Independent testing will matter more once the weights are public.

The pricing shift

The more concrete change for enterprise buyers is cost. Preview API pricing is listed at $0.68 per million input tokens, $0.07 per million cached input tokens, and $2.09 per million output tokens. 4 eWeek compared the two generations and found Large 4's API rates higher than Large 3's, even as the model gained context length and agent, coding, and security capabilities. 3

This creates some friction with Mistral's message. For years the company's appeal rested on being efficient, open, and affordable. A bigger, more capable model that costs more per token is not unreasonable. Still, it means "sovereign" AI from Mistral now comes at a higher price for teams that depend on its hosted API.

Reading the launch

Taken together, this is a significant technical step that arrives with conditions attached. On the technical side, ML4 appears to be a real advance for Mistral: a trillion-parameter MoE trained on European hardware, with a context window four times larger and credible claims of leading open-weight development outside China. 145

The open-weight part of the pitch, though, is still a promise for now. Weights are three weeks away. Early access goes first to governments and security professionals, and the version they receive has looser safeguards than what the public will get. 25 Meanwhile, the default way to use the model is a paid API that costs more than its predecessor. 3

This can be read as a reasonable choice. Releasing a trillion-parameter model to selected security testers before publishing it widely is a cautious approach, especially given the cyber capabilities Mistral is advertising. 25 What remains open is whether the October 27 release arrives on schedule and whether independent benchmarks support the claims. If both happen, ML4 strengthens Europe's position in open AI. If the release slips or the benchmarks fall short, the sovereignty argument loses much of its force.