LLM-jp-4 VL 9B
LLM-jp-4 VL 9B combines the 8B Thinking language model with Google's SigLIP 2 So400m vision encoder and a two-layer projector.
Research release
The final model replaces the beta data mixture
The final release uses Jagle, RefinedVision, selected Nemotron image data and text reasoning examples. The project removed or regenerated subsets with license or quality concerns. It also changed the language component to Thinking and produces analysis and final-answer messages.LLM-jp-4 VL 9B official model card
Parse the final answer separately
Practical guidance. A document or image application should preserve the distinction between analysis and the answer shown to the user. Compare this release with the beta using the same images, because the training and response format both changed.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Status | Research release |
| Access | Downloadable |
| Model identifiers | llm-jp-4-vl-9b |
| Components | 8.6B language model; 0.4B vision encoder; two-layer MLP projector.LLM-jp-4 VL 9B official model card |
| License | Apache 2.0.LLM-jp-4 VL 9B official model card |
| Training | About 29.3 million examples; all three components trained for 120,000 steps.LLM-jp-4 VL 9B official model card |
Related models
- LLM-jp-4 8B Thinking
LLM-jp / NII LLMC · LLM-jp-4
LLM-jp-4 8B Thinking provides Japanese and English text reasoning weights with a 65,536-token context.
Launch date not confirmedResearch releaseDownloadable - LLM-jp-4 VL 9B beta
LLM-jp / NII LLMC · LLM-jp-4 VL
The earlier VL 9B release uses the 8B Instruct backbone and dynamic image tiles.
Launch date not confirmedResearch releaseDownloadable