Janus-Pro 7B

Janus-Pro 7B is DeepSeek's unified multimodal model that handles both multimodal understanding and image generation in a single autoregressive transformer — instead of the diffusion approach used by DALL-E 3 and Stable Diffusion. Its key innovation is decoupled visual encoding: separate encoding pathways for understanding (SigLIP-L vision encoder, 384×384 input) and generation (tokenizer with ×16 downsample rate), which resolves the conflict between the encoder's two roles while keeping one unified transformer.

Benchmarks are provider-reported (DeepSeek internal testing, not independently replicated at the exact reported numbers): GenEval 80% and DPG-Bench 84.2, which DeepSeek claims beat DALL-E 3 and Stable Diffusion XL on the same tests. Press coverage at launch treated these as promising but pending independent confirmation.

Released MIT-licensed (permissive, commercial OK) — unusual for a frontier-class multimodal release and a major reason for its impact. Note: DeepSeek announced no official API; access is self-host via transformers/vLLM or community demos. Successor to Janus and Janus-Pro 1B; built on DeepSeek-LLM-7b-base. A landmark release from the R1 era, though newer unified models have since advanced the state of the art.

Provider-reported (DeepSeek internal testing): GenEval 80%, DPG-Bench 84.2 — claimed to beat DALL-E 3 and Stable Diffusion XL. Independent confirmation of the exact numbers was not established at batch time.

Independent: Launch press (gadgets360, Tech Startups, Jan 2025) reported the company's claims as promising pending third-party testing.