Mistral ships Large 4, aka le Chonk: a trillion-parameter open-weight flagship betting on sovereign AI

A trillion-parameter open-weight flagship

On October 6 the French company Mistral AI unveiled its next flagship, Mistral Large 4 (ML4), better known by its nickname le Chonk. It is the company's largest model to date: a public preview is available through the API first, while the full weights are due later this month once security testing is complete.

Architecturally ML4 is a sparse mixture-of-experts model with roughly one trillion total parameters but only about 49 billion active per token. The structure addresses a familiar problem: delivering capacity close to a very large model at an inference cost closer to a mid-sized one. For companies running their own serving clusters, the active-parameter count explains the bill far better than the headline total. Sparse mixture-of-experts is not a new idea: it splits a model into many expert sub-networks and has a router pick only a few of them for each token, so the total can grow very large while per-token compute is governed mainly by the active parameters. Most high-end models of the past two years have taken this route, differing only in expert count, routing scheme and activation ratio.

It is multimodal, working across documents, charts and images. Mistral says it topped GPT-6 Astra on one visual grounding benchmark, and that instruction following, reasoning and agentic ability all live in the same model. The company positions it as a strong research assistant that can carry a problem from the first question to the final result. Deployment is part of the pitch too: Mistral says ML4 can run in a private cloud or on-premise, so customers get weights and deployment control rather than access to a remote endpoint.

CEO Arthur Mensch was blunt with journalists: the model his company just announced is above the Chinese models on certain aspects, including cyber, and the narrative that Europe cannot compete is not true. That sentence captures Mistral's position. It faces opponents on two sides — the American closed frontier labs and the Chinese open-weight models — and its differentiation lies not in absolute scale but in a combination: open weights, European deployment, and a willingness to do security research.

Cybersecurity: treating refusals as a cost

Cybersecurity best illustrates how Mistral's route differs. On the Artificial Analysis Cyber Index, an independent evaluation of how well models fix security flaws, ML4 landed in the top five. Mistral's claim is that it leads among open-weight models — with the qualifier that this means open-weight models developed outside China.

A sharper contrast comes from Cybench: by Mistral's account, as reported by IT Pro, ML4 solved 93% of its exercises, while closed models including Claude Opus 5.5 and GPT-6 Astra scored near zero — not because they lacked the ability, but because they refused parts of the task. Cybench is a public benchmark in the cybersecurity domain whose exercises are drawn from real vulnerability scenarios and typically require a model to complete the full chain from information gathering to exploitation; on such tasks, whether a model is willing to act and whether it is able to are two different variables.

Mistral's explanation is worth recording: defending software often starts with proving that a flaw is real, exactly the kind of work safety filters in closed models can block. As threat actors increasingly jailbreak those same models to support offensive activity, defenders need systems that can match those capabilities without being constrained by the same refusals. On this route, a model's willingness to act is treated as part of its defensive value. The policy implication is considerable: if a capability is blocked by safety filters in closed models but obtainable through self-deployment in an open-weight one, then the answer to which model is safer depends on whether the question is about misuse risk or about defensive need.

The company also stresses that the model still shows a strong propensity to refuse malicious cybersecurity requests. Open weights plus self-deployment, in its framing, let organisations run these capabilities under their own policies instead of accepting a model vendor's default boundaries.

Trained from scratch on European compute

Another label attached to ML4 is trained from scratch. Mistral says it trained the model on 3,800 Grace Blackwell GPUs inside its own European data centres, with the public preview served on the same systems. A significant share of the training data was multilingual, spanning more than 160 languages including every official language of the European Union.

That turns sovereign AI from a slogan into an offering: Mistral provides a fully European deployment that it operates end to end, without other digital service providers in the path. For customers in regulated industries and the public sector, where the model runs, who operates it and which law applies can matter as much as its benchmark scores.

The training process itself has an industry flavour. Mistral says it trained ML4 in partnership with sectors including engineering, finance, science and the public sector, and that the model will serve as the foundation for specialised models aimed at particular industries and tasks — building general capability first, then spinning out vertical variants.

Open weights, but not yet open

One clarification is needed here. Open weights and open source are not synonyms: the former usually means only that the parameter files can be downloaded and self-hosted, while whether training data, training code and data pipelines are published depends on the releaser, and licences may carry limits on use, industry or geography. Precisely because the weights can leave the vendor's data centre, such models are harder to recall, monitor or restrict when it comes to safety governance — the same property Mistral sells as control, governance discussions treat as a new risk surface.

The boundary needs stating plainly: as of mid-October the downloadable weights had not been released, with Mistral promising them by the end of the month. Until then the only thing outsiders can touch is the API preview, so open weights are a stated commitment rather than something anyone can inspect yet.

Licensing deserves attention too. Multiple reports note that ML4's terms fall outside the standard open-source templates, so teams should read the usage conditions closely before committing. Parameter count and degree of openness are two different things.

The preview model is still moving as well. Mistral says the reinforcement learning run behind this preview is still in flight, that the model shows no signs of saturation and that substantial headroom remains; as it scales training on expanded infrastructure it expects large and rapid improvements in the weeks and months ahead. The capability on display now is a waypoint on that curve, not its end. IT Pro adds that part of the company's hope for further scaling rests on EU funding for European data centres.

Finally, judgment and caveats. This is Europe placing a clear bet at the open-weight frontier, bundling weights, deployment sovereignty and cyber capability into an enterprise pitch — consistent with the posture Mistral has long taken. But the striking numbers come from the company itself: they cannot be independently reproduced before the weights are out, and the near-zero Cybench scores of closed models largely reflect differences in refusal policy rather than differences in underlying ability.