Mistral announced the Large 4 public API preview on October 6. Its own Dense 200 image-grounding test reported 42%, compared with GPT-6 Astra's 41%, a one-percentage-point lead. One test does not establish broad model superiority.
Artificial Analysis reports a separate Intelligence Index score of 38 on different tasks. It cannot be compared directly with the 42% Dense 200 result, and neither score validates the weights due for release later.
Separate release schedules for API and weights
Using the API and downloading model weights are distinct steps. The announcement targets month-end, while Hugging Face lists October 31 as an estimate. Mistral also says reinforcement learning on the preview is continuing.
Active-parameter accounting and availability
The official documentation describes a 1.05T-total, 52B-active mixture-of-experts model with a vision encoder. Hugging Face specifies 49B active per token, or 52B when embedding and output layers are included. The figures describe different counting boundaries, not separate models.
An API evaluation today therefore does not settle the quality or operating conditions of the eventual checkpoint. Self-hosting still depends on the actual files, license and serving instructions becoming available.
The official model page lists a 1M-token context, while Artificial Analysis lists 524K. Long-input testing should check the actual limit configured for the API in use.
Mistral Large 4’s present results describe a preview service; self-hosting judgment awaits the actual weight release.