Overview
- Mistral, which announced ML4 on Tuesday, opened a paid public preview that lets developers and authorities test the model through its API.
- The company says ML4 has a total of one trillion parameters with 49 billion active parameters and was trained on 3,800 Nvidia Grace Blackwell GPUs.
- ML4 is an open‑weight model, meaning organisations can download and fine‑tune the parameters on their own servers so data and models stay inside their control.
- Mistral highlights strong results on cybersecurity capture‑the‑flag tasks and agentic workflows while acknowledging ML4 lags top closed US models on some coding benchmarks; the preview is reported to cost about €2 per 1 million output tokens.
- Mistral plans coordinated security reviews with researchers, partners and government bodies before releasing the full weights on 27 October to reduce misuse risk and to appeal to European public‑sector and enterprise buyers seeking local control.