Mistral AI Introduces Mistral Large 4 Preview, Plans to Release Weights by End of October
Mistral AI has made the multimodal Mistral Large 4 model available through a preview API. The model uses a Mixture-of-Experts architecture, but its weights are not yet public.
On October 6, Mistral AI introduced Mistral Large 4 preview, a public preview version of its new multimodal language model. The model is currently available through a preview API. The company did not release its weights when announcing it, but says it plans to do so by the end of October.
Mistral Large 4 is built on a Mixture-of-Experts (MoE) architecture. According to the technical documentation, it has approximately 1.05 trillion total parameters, while activating 49 billion parameters when processing a single token. Mistral also lists a context window with a capacity of one million tokens.
Mistral Large 4 preview is currently available only through the API
The new model is currently accessible as a preview through an API. This means developers and companies can test it through Mistral’s infrastructure, but cannot yet run it using publicly released weights of their own.
According to Mistral, the preview API costs $1.36 per million input tokens and $4.18 per million output tokens. The company describes the model as multimodal and also mentions agentic capabilities. Further technical details about the architecture, post-training, or inference requirements do not yet follow from the available information.
Mistral says it will release the weights by the end of October 2026. The Mistral organization’s page on Hugging Face lists a tentative date of October 31. Until that happens, the “open-weight” label describes a planned state, not the model’s current availability.
A large MoE model for APIs and possible self-hosted deployment
The MoE architecture divides a model into multiple expert components and activates only part of the total parameters for individual tokens. In the case of Mistral Large 4, the difference between approximately 1.05 trillion total parameters and 49 billion active parameters is significant mainly when assessing the model’s computational characteristics. However, the figures alone do not indicate the actual costs or performance in a specific deployment without independent measurements.
If Mistral releases the weights according to the announced schedule, the model could expand the range of large systems deployable on an organization’s own infrastructure. This approach is particularly relevant for enterprise and sovereign AI deployments. However, practical usability will also depend on the license, hardware requirements, and reproducible performance after the weights are released.
Benchmarks and safety claims await independent verification
When introducing the model, Mistral presented comparisons with both closed and open competing models. However, these claims are based mainly on benchmarks and evaluations published or commissioned by the company itself. Broad independent verification is not yet available.
The company also claims strong capabilities in cybersecurity and a higher rate of refusing harmful requests compared with other open models. This does not constitute confirmation of a specific incident or misuse, nor of an independently verified level of safety.
Going forward, it will be important whether Mistral meets the deadline for releasing the weights and under what licensing terms. Independent tests of coding, agentic tasks, multimodality, and cybersecurity will also be monitored, along with technical information and actual inference costs.
Sources
- Mistral AI – Confirms the release of ML4 preview, its parameters, API availability, stated benchmarks, pricing, and plan to release the weights by the end of October.
- Mistral AI Docs – Lists the model’s technical parameters, including 1.05 trillion total parameters, 49 billion active parameters, and a one-million-token context.
- Hugging Face – mistralai – The Mistral organization’s page lists the planned release of the weights and the tentative date of October 31, 2026.
- TechCrunch – Independently confirms the model’s release and that its weights were not yet publicly available at the time of the announcement.
Verified and updated: 10/07/2026 06:21
