Mistral bets on open cyber-capable AI to challenge China as US labs stay closed
Photo by Solen Feyissa on Unsplash
- Mistral has previewed Large 4, a trillion-parameter open-weight model aimed at challenging China’s GLM-5.3.
- Mistral is pitching Large 4’s cyber capabilities as an alternative to restricted American models.
- The company plans to release downloadable weights on October 27 after a three-week security test.
Mistral has launched a public preview of its trillion-parameter open-weight model, Mistral Large 4 (ML4), which it is calling the most powerful of its kind built outside China.
Cybersecurity is the leading pitch of the model that has been code-named “Le Chonk.” The model is available through Mistral’s API. In its announcement, the company stated that it will be publishing the downloadable weights around October 27 after a three-week test with developers, security firms, and government agencies.
A week earlier, on September 29, Anthropic shared findings from its research, warning that Zhipu AI’s open-weight GLM-5.3, known outside China as Z.ai, can autonomously build working cyber exploits. It also stated that it could ship with safeguards weak enough to be bypassed most of the time.
“Frontier of open weight models” and what the benchmarks show
Co-founder and chief scientist Guillaume Lample said, “ML4 is at the frontier of open weight models.”
The model, which was launched on Tuesday, October 6, has been put through a preliminary test on DeepSWE v1.1, an agentic coding test, putting Large 4 at 62%, just one percent above GLM-5.3.
Why does Anthropic’s GLM-5.3 warning matter?
Anthropic’s research helps Mistral’s positioning of its product.
In a September 17 assessment cited by Anthropic, the US government’s Center for AI Standards and Innovation called GLM-5.3 “the most cyber-capable open-weight model released to date,” trailing the US frontier by about four months.
Anthropic’s own Frontier Red Team says that it discovered that the model could chain together previously unknown browser flaws into a working exploit in under a day of lightly supervised work.
The safeguards are the problem Anthropic highlights. The company wrote that attackers bypassed GLM-5.3’s built-in protections “between 64% and 100% of the time” using simple techniques in simulated tests. Anthropic stated that the same attacks failed against safeguarded Claude models.
Open weights against the US labs’ locked models
The leading American AI labs have largely kept their most capable models out of general release. An example of that is Anthropic Claude Mythos Preview, which is its most capable cyber model, and it is not open for general release.
What Anthropic is doing instead is to route the model through Project Glasswing to about 40 vetted organizations because it judged the model as too dangerous to hand out.
OpenAI also took a similar route in April when it gated its GPT-5.4-Cyber variant behind tiered identity checks in its Trusted Access for Cyber program. In September, it said that its Astra model crossed the “Critical” cyber threshold in its Preparedness Framework.
Mistral is taking a different route from its American counterparts. It is not gating the model, as it says, arguing that security teams need to scan their own systems daily. These teams also see that closed models are often too costly at that volume and, in some instances, refuse dual-use requests outright.
“The cyber defense capabilities will enable enterprises and governments to defend themselves against threat actors that are jailbreaking closed models to perform cyber attacks,” Lample said. Developers, cybersecurity firms, and government agencies in the preview will get a version with fewer restrictions than the public API.
A sovereignty pitch backed by a €3 billion raise
In September, Mistral announced that it raised €3 billion ($3.5 billion) at a valuation above €21 billion. Large 4 is the first model it has released since that announcement.
The model was trained from scratch over the span of two months on about 4,000 Nvidia Grace Blackwell GPUs in Mistral’s European data centers.
Don’t just read crypto news. Understand it. Subscribe to our newsletter. It's free.
FAQs
What is Mistral Large 4 and when can people download it?
Large 4, code-named "Le Chonk," is a trillion-parameter multimodal model with 49 billion active parameters, available now through Mistral's API, with the open weights set to publish on October 27 after a three-week testing period.
How does Large 4 compare to China's GLM-5.3?
On Mistral's preliminary DeepSWE v1.1 coding benchmark it scores 62% to GLM-5.3's 61%, though VentureBeat notes the live leaderboard puts GLM-5.3 nearer 69% under its best configuration, so Large 4 looks competitive among open models without an outright lead.
Why is cybersecurity central to the launch?
Mistral argues security teams need to scan systems daily and that closed models are often too expensive or refuse dual-use requests, so an open model buyers can run themselves gives them more control, a pitch sharpened by Anthropic's finding that GLM-5.3's safeguards can be bypassed 64% to 100% of the time.
Hannah Collymore
Hannah is a writer and editor with nearly a decade of blog writing and event reporting experience in the crypto space. At Cryptopolitan, Hannah contributes to the news page, reporting and analyzing the latest developments in DeFi, RWA, crypto regulation, AI and frontier tech industries. She graduated from Arcadia university with a degree in Business Administration.
