Mistral launches Le Chonk, its largest model⁠.

Mistral has launched a public preview of Mistral Large 4, which it calls ML4 and, unofficially, le Chonk.¹ The preview is open now through Mistral's own API, and the company says the weights will be released by the end of the month.¹ It describes the model as natively multimodal, with one trillion parameters, of which 49 billion are active at any one time.¹

The claim is about rank. Mistral says ML4 matches the strongest open models anywhere and clearly beats any open weight model developed in the US or Europe.¹ Wired reports that the company presents it as by far the most capable open weight model built outside China, and "very, very close" to some closed models.² Mistral says it trained the model from scratch on 3,800 Nvidia Grace Blackwell chips in its own data centres in Europe, which Wired sets against US claims that Chinese labs have copied the outputs of larger models to catch up.¹²

Security is the pitch. Mistral says ML4 ranks in the top five models worldwide on the Artificial Analysis Cyber Index, and scores 82% on one test that asks a model to reproduce a real flaw in open source software and then patch it.¹ It says several closed models, including Claude Opus 5.5 and GPT-6 Astra, score near zero on that test because they refuse the task.¹ Until the weights are out, it is testing the model with security firms, vetted partners and state authorities, who get a version with lighter moderation and wider cyber abilities.¹

The business case is control. Wired notes that open weight models cost only as much as the computing they use, and that Mistral raised $3.3 billion in September at a $24 billion valuation.² Mistral says ML4 can run on a private cloud or on premises, and offers a European deployment it runs itself under European law.¹ Its cofounder, Guillaume Lample, told Wired: "If you use a closed model, there is no guarantee it will still be there tomorrow."²

There is a British thread. Wired reports that the White House has reportedly asked the American labs to hold back unreleased models even from the UK's AI Safety Institute, after the US placed temporary limits on the spread of OpenAI and Anthropic models in June.² Lample said the race was the main message: "Mistral is still in the race of getting the best model."²

A credible choice, not yet proved

Read this as a sales document first. Almost every number in it is Mistral's, on tests Mistral chose, and the weights that make it open are still weeks away. The cyber result is partly a comparison with models that decline the task by design, which says as much about their makers' rules as about ability. That does not make it hollow. A model good enough to sit near the top, that you can run on your own kit and that no foreign government can switch off, is a real option for a British firm that worries about where its data goes.

None of the reporting names a UK customer or a price in pounds. If you rely on one American model today, the useful question is simple: what would you do if it changed terms or vanished next month? A second supplier you have tested is cheap insurance. ML4 may be that supplier, once the weights land and someone other than Mistral has measured it.

Also today

• Google moves free Gemini users to its smallest model

  From 9 October, free Gemini users get only Flash Lite, Flash needs the $4.99 a month AI Plus plan, and Pro moves up to the $19.99 Pro tier.³

• Drax data centre plan would out-emit Gatwick's flights

  A campaign group's analysis says Drax's planned North Yorkshire data centre would burn 4.9m tonnes of wood a year, nearly double the CO2 of every Gatwick flight.⁴

• Italy's prime minister files to trademark her voice

  Giorgia Meloni has applied to the EU Intellectual Property Office to trademark a four-second recording of her voice, to guard against AI deepfakes.⁵

• Misused AI tops a survey of reputation risks

  A quarterly survey of more than 150 public affairs leaders ranks misuse of AI as the top threat to company reputations, ahead of harm to children.⁶

• People paid female AI agents less than male ones

  In a University of Limerick study of 189 people, identical AI assistants were paid roughly 10 per cent less when presented as female.⁷