What it is
DeepSeek-OCR 2 is an open document-recognition model from DeepSeek, published with a paper on January 28, 2026 under Apache 2.0. It converts page images into text and Markdown, including tables and layout. Its new DeepEncoder V2, built on Qwen2-0.5B, reorders the visual pieces of a page by meaning, rather than reading strictly left to right and top to bottom, before passing them to a 3-billion-parameter mixture-of-experts decoder with about 500 million active parameters.
Each page uses between 256 and 1,120 visual tokens, far fewer than general vision models. It succeeds DeepSeek-OCR from October 2025, whose compression approach it keeps. The GitHub repository has no tagged releases and no commits since February 3, 2026.
Our review
DeepSeek-OCR 2 is a clever, efficient document reader that has been overtaken. DeepSeek's paper reports 91.09 on OmniDocBench v1.5, 3.73 points above the original DeepSeek-OCR, and cuts reading-order edit distance from 0.085 to 0.057. In the same table the pipeline system PaddleOCR-VL scored 92.86, and its successor PaddleOCR-VL-1.6 now reports 96.33 on v1.6.
Independent reruns land lower still. Baidu's Unlimited-OCR paper measures DeepSeek-OCR 2 at 89.17 on v1.5 and 90.25 on v1.6, against 93.23 and 93.92 for its own model, which is itself built on DeepSeek-OCR. The Hugging Face card adds a ParseBench mean of 41.2 and an olmOCR-bench score of 76.3 taken from a post on X. The spread shows how much OCR rankings depend on who runs the test and with which settings.
The efficiency case still holds. A page costs 256 to 1,120 visual tokens, where Qwen3-VL-235B and several end-to-end parsers in DeepSeek's table use more than 6,000, and the decoder activates only about 500 million parameters per token. The BF16 checkpoint is 6.78 GB. DeepSeek publishes no VRAM or pages-per-second figure, only that its PDF mode runs at speeds on par with the original.
Operationally it is frozen in early 2026. The reference environment pins CUDA 11.8, torch 2.6.0, transformers 4.46.3 and a vLLM 0.8.5 wheel; loading needs trust_remote_code, which executes Python from the model repository; and the GitHub project has had no commits since 3 February. A team starting a document pipeline now would be adopting a stack its maintainers have stopped updating.
Where it earns its place
A data engineer digitising a few million pages of scanned archives
Convert scans to Markdown cheaply on owned GPUs
The 1,120-token cap per page keeps decoder compute and context small, and the vLLM scripts process PDFs concurrently. Budget time to build the CUDA 11.8 and vLLM 0.8.5 environment, and spot-check tables, where newer rivals score higher.
An ML team already serving DeepSeek-OCR in production
Improve reading order on multi-column layouts without changing stack
The prompts and serving path carry over, and DeepSeek reports reading-order error falling from 0.085 to 0.057. It is the cheapest upgrade they have, though moving to PaddleOCR-VL-1.6 buys more accuracy for more migration work.
A compliance lead at a mid-sized bank evaluating document AI
Find an OCR model that runs inside the bank's network under a clean licence
Apache 2.0 with no hosted dependency fits that brief, but trust_remote_code means the repository's Python has to be reviewed and pinned to a commit before it runs on internal servers.
Before you commit
- Setup
- Transformers path: torch 2.6.0, transformers 4.46.3 and flash-attn 2.7.3 on CUDA 11.8, then model.infer with base_size 1024 and image_size 768. The vLLM path uses a vLLM 0.8.5+cu118 wheel and scripts from the GitHub repo.
- Integrations
- Two prompt modes: "<|grounding|>Convert the document to markdown." for layout-aware Markdown and "Free OCR." for plain text. No Hugging Face inference provider hosts it.
- Data and privacy
- Apache 2.0 and self-hosted, so documents stay on your hardware; loading requires trust_remote_code, which runs repository code.
- Cost at scale
- No hosted price exists; cost is GPU time. Weights are 6.78 GB and pages use 256 to 1,120 visual tokens, but DeepSeek publishes no VRAM or throughput figure to size a fleet.
- Limits
- Trails PaddleOCR-VL-1.6 and Unlimited-OCR on OmniDocBench, is tied to an old CUDA and vLLM stack, and has had no repository activity since 3 February 2026.
What's good
- Apache 2.0 with no usage restrictions
- 256 to 1,120 visual tokens per page
- Reading-order error cut to 0.057 from 0.085
- 6.78 GB of BF16 weights
What's not
- Trails PaddleOCR-VL-1.6 and Unlimited-OCR on OmniDocBench
- Baidu's rerun scores it 1.9 points below DeepSeek's figure
- Pinned to CUDA 11.8 and vLLM 0.8.5
- No commits since February 2026
Best for
Teams already invested in the DeepSeek-OCR stack who want a drop-in accuracy upgrade.
Alternatives
DeepSeek-OCR 2: common questions
- Is DeepSeek-OCR 2 worth it?
- Frugal with visual tokens and well documented in its paper, but its rivals read better and its own repository stopped moving in February. We rate it 3.7 out of 5.
- How much does DeepSeek-OCR 2 cost?
- Free under Apache 2.0.
- What are the downsides of DeepSeek-OCR 2?
- Trails PaddleOCR-VL-1.6 and Unlimited-OCR on OmniDocBench; baidu's rerun scores it 1.9 points below DeepSeek's figure; pinned to CUDA 11.8 and vLLM 0.8.5; no commits since February 2026.
- Who is DeepSeek-OCR 2 best for?
- Teams already invested in the DeepSeek-OCR stack who want a drop-in accuracy upgrade.
- How much does DeepSeek-OCR 2 cost for a team or in production?
- No hosted price exists; cost is GPU time. Weights are 6.78 GB and pages use 256 to 1,120 visual tokens, but DeepSeek publishes no VRAM or throughput figure to size a fleet.
- Is DeepSeek-OCR 2 safe for company data?
- Apache 2.0 and self-hosted, so documents stay on your hardware; loading requires trust_remote_code, which runs repository code.
- What are the best DeepSeek-OCR 2 alternatives?
- The closest alternatives we have reviewed are PaddleOCR-VL-1.6 and Unlimited-OCR.
Answers checked Oct 22, 2026.