Salamandra 7B Vision
Salamandra 7B Vision is BSC's earlier image- and video-understanding research model. Its card now marks it as superseded.
Research release
Model and training
The model combines Salamandra 7B Instruct with Google's SigLIP SO400M patch-14 image encoder and a newly trained two-layer projector. BSC used the LLaVA OneVision framework for multimodal instruction tuning. It generates text from visual input, rather than generating images. The card recommends Salamandra-VL-7B-2512 instead and identifies the license as research-only.Salamandra 7B Vision: official model card
Practical use
Practical guidance. Keep this model for reproducing earlier visual-language results. Its third-party image encoder and BSC language backbone should both be recorded in provenance. For new research, compare the successor on document images and multilingual questions. Do not interpret an image-understanding checkpoint as a video-generation model.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Status | Research release |
| Access | Downloadable |
| Model identifiers | salamandra-7b-vision |
| License | RESEARCH-ONLY RAIL-AMS.Salamandra 7B Vision: official model card |
| Model size | 7B language backbone plus visual encoder and projector.Salamandra 7B Vision: official model card |
Related models
The original Salamandra 7B Instruct is BSC's first conversational adaptation of its 7B multilingual base.
Launch date not confirmedResearch releaseDownloadable- Salamandra VL 7B 2512
Barcelona Supercomputing Center · Salamandra Vision
Salamandra VL 7B 2512 updates BSC's visual-language model with a new image encoder and additional multilingual training.
Launch date not confirmedResearch releaseDownloadable