Mistral releases trillion-parameter model firms can run themselves
Mistral, a French AI company, published the weights of its Large 4 model with one trillion parameters on 29 October, aimed at software development and cyber defence.
For the first time a European model of this size is freely downloadable for companies and public bodies to run on their own servers. That reduces dependence on American providers where it is most sensitive: security.
What happened
According to its maker, Mistral Large 4 has one trillion parameters, is multimodal and was trained on European infrastructure. The weights are on Hugging Face; the licence permits use, modification and commercial deployment, and only providers of AI as a service need a separate licence. Through the API, one million input tokens cost $0.70 and one million output tokens $2.10. Mistral is marketing the model for sovereign cyber defence, according to TechCrunch: it is meant to detect threats in code and IT systems and can be run on-premises. On a cyber index by Artificial Analysis it scores 51 per cent, according to the company, and it refuses 96.6 per cent of malicious cyber requests. Mistral offers a dedicated security variant only to vetted partners. Assessment: the benchmarks come mostly from the company and have not yet been independently confirmed. On the same day CrowdStrike showed how attackers use freely available models for espionage. Open models strengthen defenders and attackers alike.
What follows
Companies in the EU gain an option to run large AI models under their own control without handing data to American clouds. That matters for banks, insurers and public bodies under strict data-protection rules. The cost: a model of this size requires its own data centres or expensive graphics chips. For Mistral the release is a bet that income from services and customisation will make up for giving the model away. If it works, Europe moves from buyer to supplier of a key technology. The contrast is plain on the same day: while Audi cuts its targets, the value of the new economy is created in data centres and software.
What changes in practice
- Companies
- A large AI model can run on their own servers without sending data to American clouds. Expensive computing hardware is required.
- IT security teams
- The model is built to find and fix vulnerabilities. The company's figures have not yet been independently checked.
- Public authorities
- Open weights mean access cannot be switched off. That also applies to attackers.
What to watch
- Independent tests of Large 4, for example by ENISA, the EU cyber-security agency
- First deployments at European banks or public authorities
- Cases of misuse involving the open model
Sources
- Mistral AI: Mistral Large 4
- TechCrunch: Mistral releases weights for its trillion-parameter AI model targeting cybersecurity