Janus Pro 7B

Sobre DeepSeek

Fundado em 2023 · Hangzhou, China

DeepSeek (深度求索) is a Chinese AI research lab founded in May 2023 in Hangzhou by Liang Wenfeng, the founder of the quant hedge-fund High-Flyer (which funds the lab). DeepSeek became globally prominent in late 2024 and early 2025 with the release of DeepSeek-V3 (Dec 2024), DeepSeek-R1 (Jan 2025) and the Janus multimodal family. The Janus series is DeepSeek's unified understanding-and-generation model: Janus (Oct 2024), JanusFlow (Nov 2024) and Janus Pro (Jan 2025) extend a single Transformer to both interpret images and generate them with a decoupled visual encoder design. All Janus models are open-sourced under a permissive license on Hugging Face and GitHub.

Arquitetura

Unified autoregressive multimodal Transformer (understanding + image generation)

Janus-Pro-7B (January 2025) is the largest member of the Janus family from DeepSeek. The model unifies multimodal understanding and image generation in a single autoregressive Transformer, but uniquely decouples the visual encoders for the two tasks: a SigLIP-style ViT encoder is used for image understanding (so visual features are semantic), while a VQ tokeniser based on LlamaGen is used for image generation (so visual features are reconstruction-oriented). Both feature paths are projected to the same 7B LLM backbone, which generates either text or image tokens depending on the task. For image generation Janus-Pro produces 384x384 images by autoregressive sampling of VQ tokens which are then decoded by the LlamaGen decoder. Janus-Pro improves over Janus by scaling data to ~90M image-text pairs and adding a second stage of supervised fine-tuning. DeepSeek reports that Janus-Pro-7B beats DALL-E 3, SD3-Medium and SDXL on GenEval and DPG-Bench, despite being a much smaller and unified model.

Parâmetros

7B parameters (Janus-Pro-7B)

Contexto

4.096 tokens

Capacidades

• Unified image understanding + image generation in one 7B model
• Open weights under permissive DeepSeek license
• Outperforms DALL-E 3 and SD3-Medium on GenEval (per DeepSeek paper)
• 384x384 native generation resolution
• Compatible with Hugging Face Transformers and vLLM
• Useful for multimodal agents that both see and draw
• Strong instruction-following thanks to LLM-style backbone
• Best for: research, multimodal agents, prototyping unified pipelines, fine-tuning.

Treinamento & licença

Pretrained on a mix of ~90M image-text pairs, text-only data and image-only data. Janus-Pro-7B adds extra supervised fine-tuning stages and a larger unified dataset compared to Janus 1B.

Licença: DeepSeek Janus License — open weights, free for research and commercial use with attribution and standard restrictions.

Testes de segurança: DeepSeek applies content filters and reports safety evaluations in the Janus-Pro technical report, but the open weights ship without integrated safety classifier.