LLM-jp-4 VL 9B beta
LLM-jp-4 VL 9B beta combines LLM-jp-4 8B Instruct, SigLIP 2 So400m and a lightweight projector. Its architecture draws on InternVL3.
Research release
Dynamic tiles preserve more image detail
The beta divides images into tiles based on their aspect ratio. It encodes the tiles separately before passing image tokens to the language model. The text component is the SFT-only Instruct model, not the later final release's Thinking model.LLM-jp-4 VL 9B beta official model card
Keep beta results tied to beta inputs
Practical guidance. Preserve image tiling settings when reproducing beta evaluations. Fine-detail results can change when preprocessing changes, even if the same model files are used.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Status | Research release |
| Access | Downloadable |
| Model identifiers | llm-jp-4-vl-9b-beta |
| Training data | 24.2 million FineVision and 9.2 million Jagle examples.LLM-jp-4 VL 9B beta official model card |
| Training | 90,000 steps; approximately 180 billion tokens including image tokens.LLM-jp-4 VL 9B beta official model card |
| License | Apache 2.0 is declared for the weights. Check the source-data terms.LLM-jp-4 VL 9B beta official model card |
Related models
- LLM-jp-4 8B Instruct
LLM-jp / NII LLMC · LLM-jp-4
Japanese-English instruction model trained with SFT alone, distinct from the 8B Thinking release.
Launch date not confirmedResearch releaseDownloadable - LLM-jp-4 VL 9B
LLM-jp / NII LLMC · LLM-jp-4 VL
A Japanese-English vision-language reasoning model that changes both the language backbone and training data from the beta.
Launch date not confirmedResearch releaseDownloadable