Janus-Pro

Janus-Pro is DeepSeek's open-weight autoregressive multimodal model. It decouples visual encoding, using a SigLIP-L encoder for image understanding and a separate VQ tokenizer for generation, so one model handles both tasks. It comes in 1B and 7B sizes; the 7B was reported to beat DALL-E 3 and Stable Diffusion 3 Medium on text-to-image instruction-following benchmarks at release. Honest limitations: it is an early-2025 model, and newer open multimodal models have moved past it. A community reproducibility report on GitHub found much lower MMBench scores (around 66 vs the reported 79.2) under a different evaluation harness, so the provider's understanding benchmarks should be read with that caveat. DeepSeek has not announced a hosted API, so use means self-hosting with a strong GPU.

Provider-reported (Jan 2025 paper, arXiv 2501.17811): GenEval 0.80 for the 7B model (vs DALL-E 3 at 0.67, SD3 Medium at 0.74). DPG-Bench 84.2. MMBench 79.2 for multimodal understanding (vs Janus 69.4, TokenFlow 68.9, MetaMorph 75.2). Caveat: a GitHub issue reports community reproduction of MMBench-EN at ~66 with lmms-eval, suggesting the reported figure depends on evaluation protocol.