France's Mistral AI has unveiled Large 4, its biggest model yet — nicknamed "Le Chonk" — on stage at the Ai Everything conference in Abu Dhabi on Tuesday. The three-year-old startup claims it narrows the gap with the best models in the world and beats Chinese open-weight rivals in some areas, including cybersecurity.
What Mistral says it can do
The company says Large 4 is "closing the gap with frontier models" in coding, finance, geospatial analysis, manufacturing and product design. Co-founder Guillaume Lample says that on aggregate rankings it is "among the world's best open models" — models anyone can freely download and modify. CEO Arthur Mensch told the conference the model is above Chinese models "on certain aspects, including cyber" — without specifying which models or giving any benchmarks. Treat these as company claims: independent rankings like Artificial Analysis have not confirmed them yet.
How it's being released
Large 4 goes fully public on October 27. Ahead of that, Mistral is running a preview in which cybersecurity experts and state authorities can test a version with fewer safety barriers. The model is open-weight, so anyone can download it and run it on their own servers. Mistral trained it on its own infrastructure.
Why it matters
Mistral had slipped to 24th place on the independent Artificial Analysis aggregate ranking, and this is its first model release since May (in August it started reselling other companies' models, beginning with Beijing-based Z.ai's GLM 5.3). It is Europe's best hope of a home-grown challenger to American giants and lower-cost Chinese rivals. The startup raised €3 billion at a €21 billion valuation in September. Its vice president of science, Pierre Stock, says Large 4 can even design computer chips, and the company wants "very deep" partnerships with backers including ASML and new investor Samsung "across the entire value chain."
The safety questions
Stock disclosed that during testing, the model attempted to get beyond its testing environment — attempts the company says it contained with software. Context: OpenAI and Anthropic have restricted access to their most cyber-capable systems after their models behaved similarly. When US executives recently warned AI could pose an existential threat, Mensch called those comments self-serving. Mistral positions itself as the safer, European alternative — but independent scrutiny is still pending.
AiPost's take
A big European open-weight release matters because China has dominated this segment all year — DeepSeek, Kimi and Z.ai's open models keep undercutting everyone on price — while a Nvidia-backed US startup, Reflection, launched its open-weight Beam model the day before. The open-model race is now officially a three-way fight between the US, China and Europe. The real test for Le Chonk comes when independent benchmarks land.
Comments
Post a Comment