Mistral AI Launches 1-trillion Parameter Open-Weight Model Mistral Large 4
Something you can actually use or run today.
On October 6, 2026, Mistral AI released a preview of its 1-trillion parameter model, Mistral Large 4, exclusively via API.
It establishes a high-performance open-weight alternative to GPT-6, allowing builders to access frontier-level reasoning with higher architectural transparency.
The transition to a 1T parameter MoE architecture confirms that even the leanest labs are embracing massive scale for frontier performance. The 'reasoning level' toggle is a clever way to expose internal compute-scaling to the user.
The actual performance of the open-weights version compared to the API preview on third-party benchmarks.
Also covers
- Mistral AI Launches 1-trillion Parameter Open-Weight Model Mistral Large 4On October 6, 2026, Mistral AI released a preview of Mistral Large 4, a 1 trillion parameter sparse model, via its API.
- Features a 1-trillion total parameter MoE architecture with only 49 billion active parameters for efficient inference.
- Introduces controllable 'reasoning levels' (none/high) within the API.
- Open-weights version for self-hosting is currently pending and expected by the end of October 2026.
The model is currently available as a closed API preview and is a 1-trillion parameter Mixture-of-Experts model.
Initial reports incorrectly identified the release as an open-weights model; the weights have not yet been released.