AI
Mistral previews a trillion-parameter open-weight model aimed at coding and cyber work
Mistral has opened a preview of Large 4, a 1-trillion-parameter multimodal mixture-of-experts model with 49 billion active parameters. The company says it leads open-weight models in cybersecurity, coding and several enterprise tasks, while reporting strong results on selected external evaluations and internal tests. Mistral plans to release the model weights by month-end and supports deployment on private infrastructure. The evidence is primarily a company announcement; several benchmark scores were evaluated privately and broader architecture and training details are still forthcoming.
Editorial summary · Publisher article
What matters
- The preview is served from Mistral’s European infrastructure, with weights promised by the end of October.
- Mistral reports an 82% score on a vulnerability-reproduction and patching test and 93% on Cybench.
- The model scored 49.8% on the company’s combined Coding Agent Index and 59.9% on AutomationBench.
- Mistral says Large 4 was trained on 3,800 NVIDIA Grace Blackwell GPUs and supports more than 160 languages.
Context & caveats
- Performance claims largely come from Mistral or selected evaluators; some benchmark results were assessed privately before public harnesses launched.
- The weights, architecture details and post-training methodology were not yet available in the supplied evidence.
Why this made the edition
Substantive first-party launch account with model scale, deployment plans, benchmark results, infrastructure details and safety caveats. Most performance claims remain Mistral’s own or rely on privately evaluated benchmarks, so the summary preserves attribution.
Source material