DeepSeek-VL

// Overview

DeepSeek-VL is DeepSeek's first-generation, open-weight vision-language model series, released on 11 March 2024 alongside the technical report 'DeepSeek-VL: Towards Real-World Vision-Language Understanding' (arXiv:2403.05525). It comes in two dense sizes — DeepSeek-VL-1.3B and DeepSeek-VL-7B — each available as a Base model and an instruction-tuned Chat model. The design goal was practical, real-world multimodal understanding rather than benchmark chasing: the training data deliberately spans web screenshots, PDFs, OCR, charts and knowledge content, and the instruction-tuning set was built from a taxonomy of real user scenarios.

Architecturally, DeepSeek-VL pairs a dense DeepSeek-LLM language backbone (DeepSeek-LLM-1B for the 1.3B model, DeepSeek-LLM-7B for the 7B model) with a hybrid vision encoder. A SigLIP-L branch reads images at 384x384 for semantics while a SAM-B branch reads at 1024x1024 for fine detail; the two are fused into 576 visual tokens. This high-resolution path is what lets the model read documents, small text and dense charts without cropping away information, all at relatively low compute overhead. A key finding of the paper was that mixing language data into vision-language pretraining from the beginning preserves the model's text abilities — DeepSeek-VL-7B performs on par with the text-only DeepSeek-7B on language benchmarks.

DeepSeek-VL is distributed as an open-weight download: the four checkpoints (1.3b-base, 1.3b-chat, 7b-base, 7b-chat) are published on Hugging Face under the DeepSeek Model License, which permits commercial use, with the inference code released under MIT. There is no first-party per-token DeepSeek API for the VL line — it is meant to be self-hosted or run through third-party hosts such as Replicate. In December 2024 DeepSeek released the second-generation DeepSeek-VL2, which replaced the dense backbone with a Mixture-of-Experts architecture and a dynamic tiling encoder, superseding this original line.

// Benchmarks

1. MMBench (7B)73.2%
2. MMBench (1.3B)64.6%
3. SEEDBench (7B)70.4%
4. SEEDBench (1.3B)66.7%
5. MMMU (7B)36.6%
6. MMMU (1.3B)32.2%
7. MM-Vet (7B)41.5%
8. MathVista (7B)36.1%
9. POPE (7B)88.1%
10. OCRBench (7B)456
11. MMLU (7B Chat, language)52.4%
12. HellaSwag (7B Chat, language)68.4%

Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.

// Strengths

• Open weights under a commercial-use license, with MIT-licensed inference code — fully self-hostable
• Hybrid SigLIP-L + SAM-B encoder reads high-resolution images (up to 1024x1024) for documents, OCR and charts without aggressive downscaling
• Preserves language ability: DeepSeek-VL-7B matches text-only DeepSeek-7B on language benchmarks (MMLU 52.4, HellaSwag 68.4)
• Two sizes — a 1.3B variant light enough for modest GPUs and a stronger 7B variant
• Trained on deliberately real-world data (web screenshots, PDFs, OCR, charts, knowledge content) rather than only academic VQA sets
• Strong hallucination resistance for its size (POPE 88.1 for the 7B model)
• Both Base and instruction-tuned Chat checkpoints released for each size

// Best for

• General visual question answering over photos and screenshots
• Reading and answering questions about documents, PDFs and forms
• OCR and text extraction from images and dense documents
• Chart, plot and infographic understanding for analytics workflows
• Self-hosted multimodal assistants where open weights and commercial licensing are required
• On-device or modest-GPU multimodal deployment using the 1.3B variant
• Research baseline for vision-language pretraining and instruction tuning

// How to access

// DeepSeek VL — every version

The full lineage of the DeepSeek VL line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.