Mixtral 8x22B
Mixtral 8x22B provides a larger sparse mixture-of-experts model under Apache 2.0. Its pretrained and instruction-tuned checkpoints share the same family, with 39B of 141B parameters active per token.
Active parameters are not the storage size
The release supports English, French, Italian, German and Spanish and describes native function calling for the instruction model. Sparse activation reduces computation per token, but the full expert weights still need storage and an appropriate inference setup.Mixtral 8x22B release announcement
Base and Instruct publication
Mistral’s April announcement explicitly released the instruction version, while the base card links an earlier torrent. This article uses April 2024 for the family. The catalog’s combined 0.1–0.3 label is not evidence for three independent trained models.Mixtral 8x22B release announcementMixtral 8x22B primary source
Specifications & access
| Detail | Value and conditions |
|---|---|
| Launch date | April 2024 |
| Access | Downloadable |
| Model identifiers | Mixtral-8x22B-v0.1Mixtral-8x22B-Instruct-v0.1open-mixtral-8x22b |
| Context window | 64,000 tokens The model catalog advertises 64k context for this version. This is not a measured recall limit. Mixtral 8x22B model and lifecycle record |
| Weight license | Apache 2.0Mixtral 8x22B official model card |
| Access and lifecycle | Hosted API retired on 30 March 2025. Downloadable weights remain listed.Mixtral 8x22B model and lifecycle recordMixtral 8x22B official model card |
| Parameters | 141B total; 39B active per tokenMixtral 8x22B release announcement |
| Checkpoint variants | Mixtral-8x22B-v0.1 and Mixtral-8x22B-Instruct-v0.1Mixtral 8x22B official model cardMixtral 8x22B primary source |