Mistral Large 4: Europe's Open AI Models Take on China

France's Mistral AI introduced its new flagship model on Tuesday, October 6, betting that a giant open system can still compete with the closed labs of the United States and the fast-moving builders in China. The model, called Large 4 and nicknamed Le Chonk inside the company, carries one trillion parameters in total, with 49 billion active for any given token, and was trained from scratch in about two months in Mistral's own European data centers.

Co-founder Guillaume Lample said the release narrows the gap with the best models in the world, and the company described Large 4 as one of the strongest open AI models anywhere. Open AI models are systems anyone can inspect, modify and run themselves, and right now that segment is led by Chinese firms. That framing is the heart of the company's pitch: genuine openness gives buyers a control that a closed API cannot promise. Mistral is presenting Large 4 as proof that open AI models built in Europe can stay in the race.

A public preview is available now through Mistral's own interface while engineers finish security testing, and the downloadable weights are expected on October 27. Until then, the open AI models pitch is a promise rather than a finished product, since the model runs only on Mistral's machines. That is the central bet of open AI models: buyers trade a little polish today for portability tomorrow.

A French answer to the US and China

Mistral needed a win. Its last major release came in December 2025, with a smaller update in April 2026, and in the meantime its systems slipped to 24th in the aggregate intelligence ranking kept by Artificial Analysis, an independent index. American labs such as OpenAI, Anthropic and Google lead that table, followed by Chinese names including Moonshot.AI, Alibaba and Z.AI.

The timing matters because Mistral spent the summer of 2026 answering questions about whether it was still trying. After announcing large investments in its own European data centers, where it even plans to host some open Chinese models, the startup faced suspicion that it was quietly stepping out of the frontier race. Chief executive Arthur Mensch told Le Monde in a September 24 interview that pushing model capability remains the company's central goal, and Large 4 is the evidence.

Training scale tells the story of a smaller shop trying to punch above its weight. Lample said the model used two to three times fewer chips than Chinese startups can access, and a fraction of the hundreds of thousands deployed by American leaders. The training run used four thousand chips and drew roughly ten megawatts of power; Mistral wants data centers totaling one gigawatt by 2030, capacity that would be shared between training its own models and serving clients. It is an argument that open AI models can narrow a resource gap through efficiency rather than scale alone.

What Large 4 can do

On the numbers Mistral chose to publish, Large 4 looks strongest at technical work. It scored 63 percent on Deep SWE 1.1, a test of long coding tasks, putting it on par with Z.AI's much-discussed GLM 5.3 and 12th overall in a table where the top systems reach 74 percent. The company also claims the model matches the best open AI models for specialized tasks: working with large spreadsheets in finance, defending against cyberattacks, and reading satellite imagery to assess disaster damage. In industrial design and production, Mistral named Airbus and BMW as customers.

Cybersecurity is the headline claim. Mistral says Large 4 ranks in the top five on Artificial Analysis's Cyber Index, and on one test that asks a model to reproduce a real software flaw and then fix it, the model reached 82 percent. Several well-known closed models scored near zero on that test, the company added, because their safety policies make them refuse the task, which says more about policy than ability. Mensch, speaking at a conference in Abu Dhabi, said Large 4 beats Chinese rivals on cyber tasks, according to Reuters reporting cited by industry press.

The context window is another selling point: one-million-token input, aimed at legal documents, technical drawings and other long materials. Whether open AI models win on benchmarks like these will be judged by independent testers once the weights land, since nearly all the published scores so far come from Mistral itself.

The open weights bet

The part of the story that matters most for buyers is control. Mistral's head of science, Pierre Stock, said open AI models guarantee that access can never be switched off by the provider. The remark points at recent disputes in which leading American models, including Anthropic's Mythos and Fable, were temporarily restricted by the White House.

European governments and companies are the target audience. Corporate clients can adapt the weights to their own needs and run them on their own servers, an argument that lands differently in a market worried about dependence on foreign AI infrastructure. Until the release, the preview runs through a monitored interface, and eventually Large 4 becomes the default model for Vibe, Mistral's assistant, though users will still be able to choose alternatives.

The launch lands in a busy month for AI infrastructure. Sierra and Meta published a personal agent protocol for AI commerce, and Constructor added agentic checkout to AI shopping agents, two signs of how fast the industry is moving to put systems like this to work.

What comes next

Mistral says it is not done. The company described itself as on a trajectory to keep improving, with Large 5 and 6 planned over the coming year on the same technical architecture. The chips to do it should come online over the next six months, funded in part by the startup's recent fundraising rounds, which included three billion euros in early September.

The skepticism that greeted the data center announcement will not vanish with one model. Lample himself put Large 4 at essentially the same level as the best Chinese systems from a month or two ago, which means it is catching up to where rivals were, not passing them. Still, for European buyers shopping for open AI models, October gave them a new option built on the continent, trained on the continent, and meant to stay open. Full specifications are detailed in industry coverage of the launch. For teams evaluating open AI models, the weight release is the real test: independent rankings will show where Large 4 sits once anyone can run it.