On October 6, 2026, Paris-based Mistral AI launched an open preview of its largest model — Mistral Large 4 ("Le Chonk"). The model, built on a multimodal mixture-of-experts architecture, packs one trillion parameters in total, but only 49 billion of them activate during inference. The company plans to release the weights on October 27, 2026 under a custom license — until then, the model will spend three weeks being tested by developers, cybersecurity leaders, and government agencies.
One Trillion Parameters, but Only 49 Billion Active
ML4 is built on a sparse architecture: while the model's total capacity is one trillion parameters, only a small fraction of the network — 49 billion parameters — fires for each request. This continues a growing trend among ultra-large open models: total capacity grows while the active share stays small.
According to Mistral, the model was trained from scratch in roughly two months — using 4,000 Nvidia Grace Blackwell GPUs housed in the company's own European data centers. For comparison, the previous Large 3 (675 billion parameters, 41 billion active) was trained on 3,000 Nvidia H200s. The company stresses that the hardware required for training is modest compared to the resources of the largest American labs, though that comparison is hard to verify precisely — competitors do not disclose training compute in a uniform format.
The model was trained on more than 160 languages, including all official EU languages. It accepts multimodal inputs but produces text output — the company positions it for software engineering, cyber defense, financial analysis, satellite and aerial imagery, technical drawings, and chip design.
Benchmarks: Mistral's Numbers Are Strong, Independent Verification Is Not Yet In
Mistral claims Large 4 is the most capable open-weights model developed outside China and can compete with the best Chinese open systems. The preliminary results it published: 62% on DeepSWE v1.1, a long-horizon software engineering benchmark; 15% on Harvey's Legal Agent Benchmark; and 93% on the Cybench cybersecurity benchmark.
VentureBeat cautions, however, that these figures deserve scrutiny. DeepSWE's live leaderboard, which selects the best published configuration per model, shows GLM-5.3 at roughly 69% — meaning ML4's 62% preview score, while competitive, especially against Western open models like Beam, does not prove outright leadership among all existing configurations.
"ML4 is at the forefront of open-weights models," Mistral co-founder and chief scientist Guillaume Lample told VentureBeat. He said the model's capabilities should improve further once reinforcement learning is complete and compute capacity expands.
Independent evaluations are not yet available: at press time, ML4 appeared neither in Artificial Analysis' open evaluations nor in the DeepSWE leaderboard. So the company's claim of being "the strongest outside China" remains provisional until independent experts test the final model and the released weights. Ars Technica reached the same conclusion.
Weights on October 27, 2026 — After Three Weeks of Testing
The company plans to serve the model through its Mistral API and publish the weights after roughly three weeks of testing — on October 27, 2026. During that period, developers, cybersecurity leaders, and government bodies will vet the model; Mistral intends to continue reinforcement learning and tune the final checkpoint in the meantime.
The weights will ship under the company's custom license. API pricing has not been announced. The central idea of the staged release strategy is to present ML4 not merely as a general-purpose model but as a foundation that enterprises and governments can run and customize themselves — including on sovereign infrastructure and in zero-data-retention mode.
From the "Le Chaton Fat" Meme to a Real Flagship
The "Le Chonk" nickname is no accident — it continues a joke that took over the internet in the summer. In June 2026, a meme about a fictional French model called "Le Chaton Fat" spread across X and Reddit, complete with fake benchmark charts and increasingly absurd specs — supposedly a giant model outclassing American and Chinese rivals. Some versions claimed over 30 trillion parameters and a stated capability of "1000 meows per second." Mistral CEO Arthur Mensch joined in the joke, writing on X: "It's actually le gros chaton."
Behind the joke was real anticipation: by July, TechCrunch was writing about speculation that Mistral was preparing a major open-weights model. Now the company genuinely has a trillion-parameter model — and it decided not to let the joke go to waste: internally, ML4 is called "Le Chonk," a nod to internet slang for an excessively large cat.
Europe's Sovereignty Bet
ML4 arrives at an inflection point for Mistral. Founded in 2023 by former DeepMind researcher Arthur Mensch and former Meta researchers Guillaume Lample and Timothée Lacroix, the company quickly became Europe's most prominent challenger to US and Chinese labs. In September, the company announced a €3 billion Series D — at a valuation above €21 billion, which it called the largest equity raise by a European tech company.
Mistral now says it supports more than 125 global enterprises, including Airbus, ASML, and HSBC. Its strategy extends far beyond weights: developer and enterprise products, customization services, inference infrastructure, and Mistral Compute. Lample told VentureBeat that customers increasingly demand not just model access but deployment, infrastructure, and surrounding engineering — the company's bet is that as weights commoditize, high-value business will migrate to the systems built around them. After October 27, 2026, what matters more than the "Le Chonk" nickname will be what developers can reproduce on their own hardware.
