Mistral's ML4 Hits Open-Weight SOTA on Security, Scores 82% on Vulnerability Patching
GuillaumeLample · x · 2026-10-06
Mistral chief scientist Guillaume Lample detailed the company's new ML4 model: SOTA among open models across many capabilities, even beating the best models in visual grounding, legal, and spreadsheet tasks.
- Security: ranks among the best on AA Cyber Index, scores 82% on vulnerability reproduction/patching and 93% on Cybench.
- Training: 3,800 NVIDIA Grace Blackwell GPUs in Mistral's European datacenter at Bruyères-le-Châtel, built with Series B funds.
- Infrastructure: Series C and D clusters coming online, enabling longer training and more ambitious post-training, with 'large and rapid improvements' expected in weeks to months.
Related event: Mistral Unveils Large 4, a 1T-Parameter Open-Weight Multimodal Model(34 posts)→
More from Infra
- Google Buys 890 MW of Nuclear Without Building a Single New Reactor — MicahBerkley · 2026-10-06
- Cycle.io Launches DevOps MCP: 3-Node Mongo Replica Set Across 3 Clouds in 15 Minutes — AlexMattoni · 2026-10-06
- Weaviate ships query profiling: a 48ms slow query turned out to be disk reads, not HNSW — CShorten30 · 2026-10-06
- HN Debate: Did Oracle Just Trigger the Implosion of the AI Bubble? — mpweiher · 2026-10-06
- Lambda adopts NVIDIA's AIPerf for model cards showing real-workload inference benchmarks — TheZachMueller · 2026-10-06
- Nvidia nears $6 trillion market value as AI frenzy keeps pushing stocks higher — AryHHAry · 2026-10-06