Janus
DeepSeekJanus is DeepSeek's family of unified models that both understand images and generate them from text in one autoregressive transformer. It separates visual encoding for the two jobs, SigLIP for understanding and a discrete image tokenizer for generation. Janus-Pro comes in 1B and 7B sizes on DeepSeek-LLM bases, and JanusFlow swaps in rectified flow for generation.
Openness
3 high confidence- weights
- open(PyTorch .bin weights on the Hub, ungated)
- data
- described(the Janus-Pro paper names the SFT additions and about 72 million synthetic aesthetic samples
- code
- partial(inference, generation and demo code
- license
- DeepSeek-Model-License(the DeepSeek License Agreement 1.0, word for word the DeepSeek-LLM LICENSE-MODEL
Janus-Pro is released under DeepSeek's model license, which allows commercial use but forbids a list of harmful uses and requires passing those limits on. Only inference code is published, and the training data is described but not released.
- https://arxiv.org/pdf/2501.17811 recorded 2026-09-26
Stage III SFT adds datasets from DeepSeek-VL2 and about 72 million synthetic aesthetic samples; no release link
- https://cdn.jsdelivr.net/gh/deepseek-ai/Janus@main/LICENSE-MODEL recorded 2026-09-26
"DEEPSEEK LICENSE AGREEMENT Version 1.0, 23 October 2023", byte-identical to the DeepSeek-LLM LICENSE-MODEL; Attachment A lists use restrictions
- https://cdn.jsdelivr.net/gh/deepseek-ai/Janus@main/README.md recorded 2026-09-26
"Commercial usage is permitted under these terms"; quick start covers inference, text-to-image generation and Gradio and FastAPI demos
- https://data.jsdelivr.com/v1/packages/gh/deepseek-ai/Janus@main?structure=flat recorded 2026-09-26
Repository file list: inference.py, generation_inference.py, demo apps and model code; no trainer or training config
- https://huggingface.co/api/models/deepseek-ai/Janus-Pro-7B?expand[]=downloads&expand[]=cardData&expand[]=gated&expand[]=createdAt&expand[]=lastModified&expand[]=siblings recorded 2026-09-26
Hub metadata for Janus-Pro-7B: gated false, PyTorch .bin shards, license mit with license_name deepseek
Adoption
2 high confidenceHugging Face downloads over the trailing 30 days, summed across DeepSeek's four Janus checkpoints.
- https://huggingface.co/api/models?author=deepseek-ai&search=Janus&limit=100 recorded 2026-09-26
Four Janus checkpoints with 21,018 downloads in the trailing 30 days; Janus-Pro-7B 10,483
Capability
1 high confidenceJanus-Pro answers questions about a single low-resolution image and also generates images from text. Its understanding side takes one image with no document or video focus, so it sits level with Florence-2; the image generation is recorded but does not raise it.
- https://arxiv.org/pdf/2501.17811 recorded 2026-09-26
Table 3: Janus-Pro-7B MMB 79.2, MMMU 41.0, MM-Vet 50.0; "80% overall accuracy on GenEval"
- https://huggingface.co/deepseek-ai/Janus-Pro-7B/raw/main/README.md recorded 2026-09-26
"unifies multimodal understanding and generation"; "uses the SigLIP-L as the vision encoder, which supports 384 x 384 image input"
Verified 2026-09-26