DeepSeek-VL2-Small

DeepSeek-VL2-Small is the mid-size variant of DeepSeek's VL2 family of mixture-of-experts vision-language models, with 2.8 billion activated parameters. It upgrades the original DeepSeek-VL with stronger visual understanding across visual question answering, OCR, document, table and chart understanding, and visual grounding.\n\nAs a sparse MoE model it reaches competitive or state-of-the-art results against dense and MoE open models while activating far fewer parameters, making it attractive for deployments that need multimodal capability on a budget. The family was open-sourced in December 2024 with weights on Hugging Face and eval code via VLMEvalKit support.\n\nIt ships under DeepSeek's model license, which permits broad research and commercial use within its terms — check the license before production deployment.