Mistral Ships a One-Trillion-Parameter Model Named 'Le Chonk,' Says Weights Will Follow in Three Weeks
Europe's flagship AI lab has a very large cat-meme model, a very large claim, and a very specific position between two other models on a leaderboard.
Mistral AI released Mistral Large 4 on Tuesday, a model nicknamed “Le Chonk.” Per TechCrunch, the name refers to its size, one trillion parameters; The Register calls it a cat-meme-inspired “very official” moniker. Mistral calls it the strongest open-weights model from the US or Europe on aggregated benchmarks, a claim with a precise geographic carve-out, in the way that “best pizza in this building” is a precise geographic carve-out.
Some precise facts first, because they matter. Large 4 is a mixture-of-experts model with roughly 1 trillion total parameters, of which 49 billion are active at a time. It is natively multimodal. It is available through Mistral’s API now, in preview. It is not yet open-weight. The weights are expected in about three weeks, after safety testing, with the company pointing to late October. The license has not been announced. So the “open” in the headline-friendly “open-weight trillion-parameter model” is, as of this morning, a forward-looking statement.
For comparison, its predecessor, Mistral Large 3 from December 2025, had 675 billion total parameters and 41 billion active. Mistral has therefore grown the model by about half in ten months, which is also, notably, roughly the rate at which one gains weight in a good French bakery.
The company’s pitch is geopolitical as much as technical. TechCrunch reports that Mistral presents Large 4 as an alternative to both closed models and the open models most often made in China, and that French president Emmanuel Macron has described the approach as “a third way in AI.” Mistral also says it trained the model on its own compute, on about 4,000 Nvidia GPUs, which Mistral VP of Science Pierre Stock described as “two to three times less than our Chinese competitors.” The Register puts the figure closer to 3,800 Grace Blackwell GPUs, about 52 NVL72 racks, in its European data centers. Either way, the number is small enough to be impressive and large enough to require an electricity bill.
Now the independent assessment. According to The Register, Artificial Analysis places the Large 4 preview between DeepSeek V4.1 Flash and OpenAI’s entry-level GPT-6 Luna on its leaderboard. It trails the flagship models from OpenAI and Anthropic, but outperforms Thinking Machines Lab’s Inkling, the most capable US open-weights model in the comparison. Mistral’s own benchmarks say it competes with top Chinese models from Alibaba, Moonshot, Z.AI and DeepSeek on coding, agentic, finance and legal work, and The Register advises treating vendor figures with caution.
Thus: one trillion parameters, and a spot between a “Flash” and a “Luna.” It is as though a heavyweight boxer entered the ring and was classified, upon inspection, as the second-best middleweight. Still, the category it wins is a real one, and in a sport this crowded a category is all anyone can really hope for.
The nickname deserves a moment. In an industry whose flagship products are named Astra, Luna and Sol, celestial names chosen to suggest infinite scale and quiet menace, Mistral looked at its largest model ever and named it after a fat cat. The Register describes it as a cat-meme-inspired moniker. This is, frankly, the most honest naming decision of the year. Every other lab says vast. Mistral says chonky. Same idea, better marketing, and nobody will ever be afraid of it.
Mistral is aiming Large 4 at cybersecurity, finance and chip design, and has said it will work with trusted partners and governments so the eventual open weights can be used for defense rather than malicious attacks. The company argues that refusals by proprietary models can get in the way of legitimate red-teaming and incident response, and says the model “pairs top-tier cyber performance with open weights and self-deployment.” That is a bold line for a model that, for the next three weeks, is behind a guardrail endpoint. Call it a trial period for the philosophy.
The competitive picture is clearer than the benchmarks. Mistral occupies an unusual position: it pitches itself as Europe’s frontier lab, a role defined partly by capability and partly by geography. It does not need to beat Anthropic or OpenAI on every chart. It needs to be good enough, large enough and open enough that European governments and enterprises feel they have an option that is neither American nor Chinese. A trillion parameters named after a chubby cat is a surprisingly effective piece of branding for that purpose.
Whether the weights arrive on schedule, and under what license, will say more about the project than any leaderboard. Until then, Le Chonk remains what every large cat is at the moment before it decides to move: massive, composed, and entirely in control of its own timeline.
Sources: