DeepSeek-VL2

DeepSeek-VL2 is an advanced series of open-source vision-language models designed for robust multimodal understanding. These models are built on an efficient Mixture-of-Experts (MoE) architecture, providing superior performance across various tasks such as visual question answering, optical character recognition, document/table/chart understanding, and visual grounding. The DeepSeek-VL2 family comprises three variants: DeepSeek-VL2-Tiny with 1.0B activated parameters, DeepSeek-VL2-Small with 2.8B activated parameters, and DeepSeek-VL2 with 4.5B activated parameters. These models achieve competitive or state-of-the-art performance with fewer activated parameters compared to existing open-source dense and MoE-based models.

DeepSeek-VL2 models are available for download via Hugging Face, and users can test them through a Gradio demo. The models support Python environments and require significant GPU memory for inference, with specific scripts provided for single and multiple image processing, as well as incremental prefilling to optimize memory usage. The models are released under the MIT License, with commercial use permitted under the DeepSeek Model License. For more detailed information, users can refer to the accompanying research paper and citation.