Janus Pro 1B
Janus Pro 1B is the smaller named checkpoint in the Janus Pro release. It keeps the family’s combined image understanding and generation design.
Research release
Compare size with observed behavior
Practical guidance. Use this checkpoint when evaluating how much model capacity your visual task needs. Compare it with 7B on the same prompts and images. A smaller download is useful only if the result meets the application’s requirements.
Image generation uses a separate representation
The model card identifies a separate image tokenizer with a downsampling factor of 16. Understanding and generation therefore do not share one identical visual encoding path.Janus Pro 1B model card
Practical guidance. Keep the supplied processor and tokenizer matched to the checkpoint. Swapping preprocessing from an unrelated vision model can change the input the model was trained to receive.
Measure the complete pipeline
Practical guidance. Measure preprocessing, model execution and image decoding together. Include the batch size and numerical precision in the result so a later deployment can reproduce the comparison.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Status | Research release |
| Launch date | |
| Access | Downloadable |
| Model identifiers | deepseek-ai/Janus-Pro-1B |
| Context window | 4,096 tokens Published sequence length. Janus project and downloads |
| Image understanding | SigLIP-L encoder with 384 × 384 inputs.Janus Pro 1B model card |
| License distinction | MIT code; model use refers to the DeepSeek Model License.Janus Pro 1B model card |
Related models
- Janus Pro 7B
DeepSeek · Janus Pro
The larger Janus Pro model combines image understanding and image generation through separate visual encoders and a shared language model.
Launch date Research releaseDownloadable - Janus 1.3B
DeepSeek · Janus
Original Janus model with separate image encoding paths; its early release required a tokenizer correction for classifier-free guidance.
Launch date October 2024Research releaseDownloadable