Janus Pro is a unified multimodal framework from DeepSeek that handles both image understanding and image generation inside a single autoregressive Transformer. Rather than forcing one visual encoder to serve two contradictory jobs, Janus decouples visual encoding into separate pathways — one tuned for perception, one for generation — while keeping a single shared transformer body. Janus-Pro, the successor to the original Janus, adds an optimized training strategy, substantially more training data, and larger model sizes, which together improve both instruction-following for text-to-image work and the stability of generated images.
This site, janusai.pro, is an independent resource hub rather than an official DeepSeek property. It collects model cards, paper links, benchmark notes, and download pointers for the Janus family — Janus-1.3B, JanusFlow-1.3B, Janus-Pro-1B and Janus-Pro-7B — and hosts lightweight in-browser demos, including a WebGPU build of the 1B model that runs locally in a modern browser without any sign-up.
The models themselves are released openly: weights are published on Hugging Face with code on GitHub, and commercial use is permitted under the stated license terms. That openness is what made Janus-Pro notable — in DeepSeek's published benchmarks the 7B model scored 0.80 on GenEval text-to-image instruction following, ahead of much larger closed systems. Because generation runs at 384x384, output is fast and cheap but noticeably softer on fine detail than diffusion-based image models.
