DeepSeek-OCR-2

DeepSeek-OCR-2 is a compact vision-language model built for one job: reading documents. Given a page image it returns plain text or a layout-aware Markdown transcript, with an optional grounding mode that ties the recognised text back to positions on the page. The weights were published on Hugging Face on 27 January 2026 and last revised on 3 February 2026; the accompanying paper, "DeepSeek-OCR 2: Visual Causal Flow", went up on arXiv the day after the release (2601.20552). The model is small by the standards of DeepSeek's language line: about 3.39 billion parameters in total, against 685 billion in DeepSeek-V3.2 and 1.6 trillion in the V4-Pro flagship. The architecture pairs a second-generation DeepEncoder vision tower with a twelve-layer decoder in the DeepSeek-V2 mixture-of-experts style — 64 routed experts plus two shared, six of them active per token — and a language context of 8,192 tokens. Images are encoded at dynamic resolution: up to six 768×768 tiles plus one 1024×1024 global view, which costs up to 6×144 plus 256 visual tokens. The whole model runs in bfloat16 on a single graphics card, under Transformers with remote code enabled or under vLLM for batch work on PDFs. The licence is the notable part. The first DeepSeek-OCR, published on 17 October 2025 as "Contexts Optical Compression", carries the MIT licence like the rest of DeepSeek's line; the second generation ships under Apache 2.0, verified in the LICENSE.txt file in the repository rather than only in the model card label. Both remain downloadable and both remain heavily used: read on 25 August 2026, the first generation records 2.37 million downloads in thirty days and the second 1.27 million, together more than DeepSeek's current flagship build. DeepSeek does not sell either model through its own API — the only way to use them is to run the weights.

▸Videos

No videos yet.

▸Reviews