Mistral AI released Mixtral 8x7B as a download link on social media, which became part of the company’s identity. It uses eight expert networks and routes each token to two of them, so it runs far cheaper than its total size suggests.
AI model data sheet: Mixtral 8x7B
| Developer | Mistral AI |
|---|---|
| Released | December 11, 2023 |
| Type | Sparse mixture-of-experts language model |
| Context window | 32K tokens |
| Access | Downloadable weights and Mistral API |
| Licence / weights | Open weights, Apache 2.0 |
| Best at | Strong quality for its cost; fast inference because only part of the model runs per token |
| Why it matters | Showed that a small European lab could ship a top open model |
Quick verdict: A landmark open model. Good starting point for self-hosted experiments.
Its permissive Apache 2.0 licence made it a favourite for self-hosting and fine-tuning.