MiniMax VL-01
VL-01 adds visual understanding to Text-01 through a 303-million-parameter vision transformer and a two-layer image projector. It is a downloadable image-and-text model.
The thumbnail preserves whole-image context
VL-01 encodes the thumbnail and non-overlapping image patches separately, then combines them. Its card reports four training stages using caption, description and instruction data.MiniMax VL-01 model card
Inspect both local detail and overall layout
Practical guidance. For document extraction, check small text and the relationship between fields. A correct number copied from the wrong table row is still an extraction failure. Compare the result with the source image before using it in a record.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable |
| Model identifiers | MiniMax-VL-01 |
| Ancestry | MiniMax Text-01 supplies the language-model component.MiniMax VL-01 model card |
| Image handling | Dynamic image grids range from 336 × 336 to 2016 × 2016. A 336 × 336 thumbnail is retained with the patches.MiniMax VL-01 model card |
| Weight terms | The card links a custom MiniMax Model License and a separate MIT code licence.MiniMax VL-01 model card |
Related models
- MiniMax M3
MiniMax · MiniMax M3
A native multimodal model with sparse attention that accepts text, images and video.
Launch date API · Downloadable · Provider product - MiniMax Text-01
MiniMax · MiniMax-01
A long-context language model that combines lightning attention, softmax attention and a mixture-of-experts network.
Launch date Downloadable