Mistral Large 4 gains downloadable weights after its API preview
Mistral released downloadable weights for an improved Mistral Large 4, extending its earlier API preview with a self-hosting option. The trillion-parameter multimodal model supports three reasoning levels and serving through vLLM or SGLang. Its custom license permits broad use but adds conditions for Model-as-a-Service businesses, with stated exceptions. Downloadable weights therefore give deployment control without making the license identical to ordinary MIT.
Mistral · Oct 29, 2026
Why it matters
Teams evaluating private deployment can now inspect the actual checkpoint, serving guidance and license instead of relying on a hosted preview.
More context
The original announcement metadata dates publication to October 29. Hosted API pricing is $0.70 per million input tokens and $2.10 per million output tokens. Review the model card’s multi-GPU serving recipe and license before a bounded evaluation.
Keep in mind
The full technical report is still forthcoming. Reported coding and security gains do not guarantee performance on local workloads, and hosted token prices do not describe self-hosting hardware costs.
Original read