Cybersecurity
Mistral releases one-trillion-parameter Large 4 preview
Mistral made a preview of Large 4 available, said the model had one trillion parameters and European training, and used benchmark scores and cyber tasks to argue for its place against closed rivals.
- Mistral released a public preview of Mistral Large 4, with the API available through Mistral Studio and model weights expected at the end of October.
- The company said the model had one trillion parameters, with 49 billion active, and had been trained in its own European data centres.
- Artificial Analysis gave the model 38 points in its Intelligence Index, above Mistral Large 3 and Mistral Medium 3.5 but below Claude Opus 5.5 at 58.
- Mistral said the model was aimed at security work and that, in one test, it scored 82 per cent where Claude Opus 5.5 and GPT-6 Astra were near zero because they refused the task.
- Mistral said it trained the model on 3,800 Nvidia Grace Blackwell GPUs and expected further improvements while the reinforcement-learning run continued.
Mistral released a public preview of Mistral Large 4, or ML4, and said the model was available through Mistral Studio. The company said weights would follow at the end of October.
Mistral described ML4 as a native multimodal model with one trillion parameters, of which 49 billion were active. Artificial Analysis gave it 38 points in its Intelligence Index, ahead of Mistral Large 3, Mistral Medium 3.5 and GLM-5.2 from Z.ai, but behind Claude Opus 5.5.
The company put IT security at the centre of the launch. It said ML4 could reproduce a vulnerability in open-source software and then patch it, while Claude Opus 5.5 and GPT-6 Astra were said to score near zero on the same test because they refused it.
Mistral said ML4 also ranked highly on refusal tests for malicious prompts and blocked 93.3 per cent of attacks in Lakera's B3 AI Security Benchmark. The company said it was training and red-teaming the model with security firms, vetted partners and government agencies until the weights were released.
Mistral said it trained the model from scratch on 3,800 Nvidia Grace Blackwell GPUs in its European data centres. It also said the training and reinforcement-learning setup would support future specialised models, and that the current run was still under way.
Named in this story
Companies
- Mistral
- released Mistral Large 4 and made the performance claims
- Z.ai
- made GLM-5.2, which Mistral said Large 4 edged past
Organisations
- Artificial Analysis
- published the Intelligence Index and Cyber Index cited in the article
- Surge AI
- ran a blind evaluation of code quality
- Vals.ai
- produced an evaluation Mistral cited on legal and financial tasks
Products and systems
- Mistral Large 4
- the new preview model at the centre of the story
- Claude Opus 5.5
- led the Intelligence Index and was cited in the cyber comparisons
- GPT-6 Astra
- was compared with Mistral Large 4 in benchmark and cyber claims
- Mistral Studio
- the service through which the preview API was available
How the source tells it
The article used a promotional, benchmark-heavy register that framed the model as a European challenge to US rivals and as a security tool.
- boosterism company claims and superlatives were repeated with little counterweight
- competitive framing leaderboard comparisons and rank changes were used to imply a major advance
- security urgency cyber work was described as something rivals would not touch, raising the stakes of the model choice
- sovereignty framing European data centres and European law were used to cast the product as a regional alternative