Pixtral 12B
Pixtral 12B combines a 12B language decoder with a 400M vision encoder. It was trained on interleaved images and text and supplies Base and instruction weights under Apache 2.0.
Deprecated API; check access details
Image dimensions affect the token sequence
The vision encoder handles different image sizes and aspect ratios. The release describes 16-by-16 image patches and separate row and image-end tokens. This preserves information about image layout rather than treating every image as an identical fixed grid.Pixtral 12B release announcement
Two checkpoint roles
Pixtral-12B-Base-2409 is the pretrained multimodal checkpoint. Pixtral-12B-2409 is the instruction version. Both support text and image input within an advertised 128K context; they produce text rather than generating images.Pixtral 12B official model cardPixtral 12B primary source
Specifications & access
| Detail | Value and conditions |
|---|---|
| Status | Deprecated API; check access details |
| Launch date | |
| Access | API, Downloadable |
| Model identifiers | Pixtral-12B-Base-2409Pixtral-12B-2409pixtral-12b-2409 |
| Context window | 128,000 tokens The model catalog advertises 128k context for this version. This is not a measured recall limit. Pixtral 12B model and lifecycle record |
| Weight license | Apache 2.0Pixtral 12B official model card |
| Access and lifecycle | Hosted API deprecated on 2 December 2025. Downloadable weights remain listed.Pixtral 12B model and lifecycle recordPixtral 12B official model card |
| Parameters | 12B decoder plus 400M vision encoderPixtral 12B official model card |