DeepSeek-OCR 2
DeepSeek's second-gen document VLM — human-like reading order at a tiny token budget.
About
DeepSeek-OCR 2 is DeepSeek's second-generation optical-compression document model, released January 27, 2026 under Apache 2.0. At roughly 3B parameters it introduces DeepEncoder V2 ('Visual Causal Flow'): instead of a plain CLIP vision tower feeding patches in fixed raster order, an LLM-style encoder reorders visual tokens by semantic reading structure before decoding — following columns, tables, and formulas the way a human reads. Typical cost is 256–1,120 visual tokens per page.
DeepSeek reports 91.09% overall on OmniDocBench v1.5 (provider-reported), a 3.73-point gain over the first generation, with particular strength in reading-order recognition. It runs on Transformers, vLLM, and Unsloth, and the paper is arXiv 2601.20552. Like its predecessor it doubles as a research artifact for long-context compression: the thesis is that rendering text as images and decoding it back can beat raw text tokens on efficiency.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
Provider-reported: OmniDocBench v1.5 91.09% overall (+3.73pp over v1), strong reading-order recognition.
Independent: no third-party reproduction found.
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 8/10
Repository verified live on Hugging Face (HTTP 200); Apache 2.0; release covered by independent press (TechNode) with consistent architecture details; paper on arXiv. Docked only for: benchmark figures are provider-run.