Token compression
A full page compresses into a fraction of the vision tokens a typical VLM needs.
DeepSeek-OCR-2
Optical context compression that packs entire pages into a handful of vision tokens.
Price$0.001 / page
ProviderDeepSeek
Parameters~3B (MoE) parameters
ApproachOptical context compression
Languages100+ languages
LicenseMIT
ReleaseDeepSeek-OCR v2 · 2026
Overview
What it does
DeepSeek-OCR-2 explores OCR as context compression: it encodes a full document image into a remarkably small set of vision tokens, then decodes faithful text. The result is extreme efficiency on dense, high-resolution pages where token budgets usually explode.
Served in base resolution mode (1024×1024, no cropping) for stable full-page parsing. Because a whole page can be represented in a fraction of the tokens a typical VLM needs, DeepSeek-OCR-2 shines on long documents and dense layouts at frontier-level accuracy.
Best for
- Long, dense document understanding
- High-resolution scans and engineering drawings
- Token-budget-constrained LLM pipelines
Benchmarks
How it measures up
olmOCR-Bench
82.4
overall accuracy score
Compression
~10×
vs. baseline VLM tokens
Image size
1024px
base parsing mode
Throughput
2.9 pg/s
on a single H100
Figures are representative of the DeepSeek-OCR-2 model card; see Hugging Face for the full evaluation suite.
Highlights
Base resolution mode
Runs at 1024×1024 without crop tiling for consistent full-page OCR.
Dense pages
Built for high-resolution, information-dense layouts and long documents.
