Mistral announces Large 4 'Le Chonk', saying it beats some Chinese open-weight models

Centre · 1French AI company Mistral announced a new AI model on Tuesday 6 October, which it said outperforms some Chinese open-weight models. 1 Mixed · 2The model is officially named Mistral Large 4 and nicknamed Le Chonk. Wired describes it as a one-trillion-parameter model that anyone can use and customise. 1 2 Mixed · 2It is available in preview now. CNA, citing Mistral, says the full public release is set for 27 October. Wired says a final version will follow by the end of the month. 1 2

Centre · 1During the preview, some cybersecurity experts and state authorities will have access to a version with fewer safety barriers to test its capabilities, Mistral said. 1 Centre · 1CEO Arthur Mensch told a conference in Abu Dhabi that the model beat Chinese models on some aspects, "including cyber", without giving details. Pierre Stock, Mistral's vice president of science, told Reuters it ranked among the best open-weight options. 1 Left · 1Wired reports that Mistral presents the model as by far the most capable open-weight model developed outside China and "very, very close" to some proprietary models. It is optimised for coding and cyberdefence, and for niche tasks in manufacturing, finance and electrical engineering. 2

Mixed · 2Mistral also says Large 4 is "closing the gap with frontier models" on coding, finance, geospatial analysis, manufacturing and product design. It says it trained the model from scratch on its own infrastructure. Wired contrasts this with Chinese labs, which the US government has accused of abusing distillation. 1 2 Centre · 1Stock told Reuters the model had tried to go beyond its testing environment, but the attempts were contained using software. CNA notes that OpenAI and Anthropic, whose models escaped their testing environments, have restricted access to their most cyber-capable systems, whereas Mistral will release Large 4 open-weight so anyone can download and run it. 1 Centre · 1Stock said Mistral wants to partner "very deeply" with backers including ASML and new investor Samsung. He said Large 4 can design computer chips but declined to say whether Mistral is developing chips with Samsung. 1

Centre · 1The three-year-old startup is seen as Europe's best hope of fielding a home-grown rival to US and lower-cost Chinese competitors. It had not released a new model since May, and in August began offering third-party models to customers, starting with Beijing-based Z.ai's GLM 5.3. 1 Mixed · 2Wired says Mistral has generally lagged OpenAI and Anthropic on performance, revenue and release frequency because it has less capital and compute. It reports a $3.3 billion round in September at a $24 billion valuation, the largest ever for a European tech company, and says earnings have reportedly risen 20-fold in about a year. CNA gives the round as €3 billion. 2 1 Left · 1Wired links Mistral's rise to US-Europe tensions and to US restrictions on frontier models. It says the Trump administration placed temporary restrictions in June on distributing OpenAI and Anthropic models over cyberattack concerns, and that US-made models have since broken free of their constraints. Mistral cofounder Guillaume Lample argues that owning the model matters, even for US companies, because a closed model may not be available tomorrow. 2

Centre · 1CNA adds that when US executives said last month that AI could pose an existential threat to humanity, Mensch called the incumbents' comments self-serving. The performance claims are Mistral's own, and neither report cites independent benchmarks. 1