Mistral AI announced public-preview access to Mistral Large 4 on October 6, 2026. The model is available through the company’s API, while downloadable weights are a separate release that Mistral planned for the end of October. Mistral AI’s announcement
Mistral Large 4 enters public preview
The announcement opened a way to use the preview through Mistral’s API. Mistral’s model documentation labels Large 4 “Public Preview” and “Open”; the company said it planned to release the downloadable weights by the end of October 2026.
Those are different stages. API access lets you send requests to a model hosted as a service. Once released, downloadable weights would give organizations the option to host the model on their own infrastructure.
What the official specifications say
Mistral’s model documentation lists 1.05 trillion total parameters and 52 billion active parameters. The first figure counts the model’s full set of parameters; the active figure describes the portion engaged during processing. They measure different aspects of the same model, not rival estimates.
| Specification | Listed value |
| Total parameters | 1.05 trillion |
| Active parameters | 52 billion |
| Vision encoder | 1.6 billion parameters |
| Architecture | Granular Mixture-of-Experts (MoE) |
| Modality | Natively multimodal, as described by Mistral |
A Mixture-of-Experts model contains multiple expert components and activates a subset for a given computation. That architecture helps explain why the total and active parameter counts differ. But parameter count describes scale; by itself, it doesn’t tell you how well a model will handle a particular task.
What downloadable weights would change
Mistral’s planned end-of-October weights release would make a different deployment option possible: organizations could run the model on their own infrastructure instead of relying solely on API access. The announcement described that release as a plan, while public-preview API access was already available.
What Mistral’s performance claims mean
Mistral reported 61.7% on DeepSWE v1.1 and 28.3% on Terminal-Bench 4. For the Dense 200 visual-grounding benchmark, the company reported 42%, compared with 41% for GPT-6-Astra. These are results Mistral published for specific evaluations; they don’t establish a universal ranking across tasks.
That distinction matters when weighing the model’s size against its results: parameter counts and benchmark scores answer different questions. A benchmark result applies to the task it measures, not every kind of work a model might be asked to do.
How Mistral says it trained Large 4
Mistral said it trained Large 4 from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its European data centers. The company also said a significant share of the training data covered more than 160 languages, including all official languages of the European Union. When it announced the preview on October 6, Mistral said the reinforcement-learning run was still in progress.
