From the archive
Mistral releases Mixtral 8x7B and opens beta API access
A mixture-of-experts model adds five-language support and a 32K context window, with downloadable and hosted routes.
Mistral released Mixtral 8x7B, a mixture-of-experts language model with weights licensed under Apache 2.0. The release includes a base model and an instruction-tuned version, alongside beta access through Mistral's mistral-small endpoint.
More capacity without using every parameter at once
Mixtral selects two of eight expert groups at each layer for each token. That lets it use part of its available model capacity for an individual computation, rather than activating the whole model each time.
The announced capabilities include:
- A 32,000-token context window.
- English, French, Italian, German, and Spanish support.
- Code generation and adaptation for instruction-following tasks.
Two ways to experiment
Downloadable weights support self-managed deployment and adaptation. The beta endpoint provides a hosted alternative, with early access registration for Mistral's generative and embedding services. The Mistral AI profile covers the company's model and API offerings.