Mistral releases Large 4 weights, with conditions for self-hosting
Mistral AI made improved weights for Large 4, its multimodal model with roughly one trillion parameters, downloadable on 29 October 2026. Businesses can now adapt it and run it on their own infrastructure.
The released version accepts text and images and offers three reasoning levels. Downloadability does not mean it runs on an ordinary computer: according to Mistral, quoted by Frandroid after a press briefing, the compressed version requires four Nvidia B200 or B300 GPUs.
The licence permits use, modification and commercial exploitation, but model-as-a-service businesses need a separate commercial agreement. Announced API prices are $0.70 per million input tokens, $0.07 for cached inputs and $2.10 for outputs.
Mistral claims improvements in coding and cyberdefence. These test results do not guarantee quality in every deployment.