DeepSeek has built a growing family of vision models that can understand and work with images. The lineup includes DeepSeek-VL, DeepSeek-VL2, the Janus series, and newer API-based vision tools. Together, they cover image captioning, visual question answering, document reading, and even image creation.
DeepSeek-VL was the first entry. It launched on March 11, 2024, as an open-weight vision-language model. It came in two dense sizes: 1.3B and 7B parameters. Both base and instruction-tuned chat versions were released. The model showed strong results on benchmarks like MMBench, MMMU, DocVQA, ChartQA, and others. It handled tasks like reading documents, answering questions about photos, and writing image captions.
DeepSeek-VL arrived in early 2024 with strong benchmark scores across document reading, visual QA, and image captioning tasks.
DeepSeek-VL2 followed on December 13, 2024. It uses a Mixture-of-Experts architecture, which means only part of the model activates for each task. It comes in three sizes: Tiny, Small, and the main VL2 model. Their activated parameter counts are 1.0B, 2.8B, and 4.5B. All three support a 4096-token sequence length. The model focuses on document, table, and chart understanding. OCR performance and visual grounding are key strengths. The full foundation behind DeepSeek-VL2 is DeepSeekMoE-27B, which contains 27 billion total parameters before the mixture-of-experts routing selects which ones activate.
The Janus family takes a different approach. Instead of just understanding images, Janus can also generate them. Janus Pro is the upgraded version with stronger performance in both understanding and creation. This sets it apart from models that only analyze visuals. Janus Pro features a unified transformer architecture that decouples visual encoding into separate pathways, enhancing its ability to handle complex multimodal operations efficiently.
DeepSeek hasn’t stopped at research models. The company now offers vision features through its API. Users can send images alongside text to get descriptions, read screenshot text, or analyze charts. In August 2026, DeepSeek announced a newer vision API called DeepSeek-V4-Flash-Vision-Exp. This points to a shift from standalone research releases toward ready-to-use multimodal services.
The pace of development has been fast. In roughly two years, DeepSeek went from its first vision model to a full product platform. The models span different sizes, architectures, and capabilities. Some focus purely on understanding images. Others can create them too. The API layer makes these tools accessible beyond the research community.
For observers tracking open-source AI, DeepSeek’s vision work represents a serious and expanding effort in multimodal technology.
References
- https://huggingface.co/deepseek-ai/deepseek-vl2
- https://huggingface.co/blog/LLMhacker/janus-pro
- https://github.com/deepseek-ai/deepseek-vl2
- https://www.deepseek.com/en/
- https://github.com/deepseek-ai/Janus
- https://blog.roboflow.com/deepseek-vision-models/
- https://cloud.baidu.com/article/3582316
- https://api-docs.deepseek.com/guides/vision/
- https://api-docs.deepseek.com/news/news260821/
- https://ai-tldr.dev/models/deepseek-vl/
