DeepSeek open source OCR-2: Re-understanding long documents with "visual causal flow"

Away from the hustle and bustle of the large model arms race, DeepSeek used a low-key open source to draw attention back to "document intelligence", the ground battlefield closest to production. DeepSeek-OCR-2 was launched on GitHub on February 3 and is open sourced under the Apache-2.0 protocol. The technical keyword is "Visual Causal Flow" - a new solution for more efficient causal modeling and compression of visual information.

From "optical compression" to "causal flow"

The technical highlight of the first generation DeepSeek-OCR is "Contexts Optical Compression", which has received approximately 23,700 stars and 2,187 forks since it was open sourced. It is currently one of the most popular OCR open source projects. The second generation evolved on its basis: emphasizing more efficient causal modeling and compression of visual information, and performing better in scenarios such as long document OCR and complex layout recognition. As of mid-2026, the warehouse has accumulated approximately 3,200 stars and continues to iterate.

Open source is not a piecemeal action, but a technology stack

Looking at DeepSeek's open source actions over the past year, OCR is just one piece of the puzzle: in terms of inference infrastructure, DeepEP (expert parallel communication), DeepGEMM (GPU BLAS kernel), FlashMLA (multi-head potential attention kernel), and 3FS (AI training/inference distributed file system) have been open sourced successively, covering the entire "algorithm-kernel-communication-storage" link. This strategy of "not only giving weight, but also giving projects" is building DeepSeek's unique trust asset in the open source community.

From an industry perspective, OCR seems to be an "old track", but in fact it is an entry-level capability for RAG, document intelligence and multi-modal implementation. DeepSeek is open sourced under the Apache-2.0 protocol rather than a looser commercial license, which not only maintains ecological influence but also leaves room for commercial closed source. For domestic teams doing document parsing, knowledge base, and bill recognition, DeepSeek-OCR-2 provides a Chinese-scenario-friendly base that can be directly commercialized—this is precisely the shortcoming of many overseas OCR models that fail in long Chinese documents.

Several directions worth tracking in the future:

1. Performance of DeepSeek-OCR-2 on the Chinese long document benchmark: The official has not disclosed the complete benchmark, and community testing is the key to judging its quality.
2. Collaboration with multi-modal models (Qwen2.5-VL, etc.): The boundary between OCR and visual language models is blurring. Will DeepSeek take advantage of the trend and launch supporting VLM.
3. Can "visual causal flow" spill over to video understanding: The extension of the same set of causal modeling ideas in the time dimension is worth observing.
4. Open source agreement and commercialization balance: Whether Apache-2.0 will introduce more stringent terms in the future will affect enterprise adoption decisions.