Mistral Launches Large 4: 1T-Param Open Model
Original: Mistral Large 4
Why This Matters
A 1T-param open-weight multimodal model shifts the ceiling for what enterprises can self-host.
Mistral released a public preview of Mistral Large 4 on Oct 6, 2026. The 1-trillion-parameter multimodal model has 49B active parameters, trained on 3,800 NVIDIA Grace Blackwell GPUs in European datacenters. Weights drop end of October.
Mistral Large 4 — internally ML4, officially nicknamed 'le Chonk' — is Mistral's largest model to date. It uses a sparse mixture-of-experts architecture: 1 trillion total parameters with 49 billion active at inference. The model is natively multimodal and was trained from scratch on Mistral's own European infrastructure using 3,800 NVIDIA Grace Blackwell GPUs.
Mistral claims ML4 is competitive with the strongest open-source models globally and outperforms any open-weight model built in the US or Europe. In visual grounding specifically, the company says it beats frontier closed models. Enterprise verticals highlighted include cybersecurity, finance, and law.
Before the weight release, Mistral is running red-teaming with cybersecurity firms, vetted partners, and state authorities — with reduced moderation and expanded cyber capabilities for those testers. The company frames open weights as a feature for security work, where losing provider access mid-incident is itself a risk.
Training data spanned more than 160 languages, covering every official EU language. ML4 shares the same training and RL environment Mistral offers enterprise customers via Mistral Forge. The preview API is live on Mistral Studio now; full weight release and architecture details are expected by end of October 2026.