OCR and documents · DeepSeek
DeepSeek-OCR
Optical context compression: renders long text as images to save tokens.
Datasheet
Maker | DeepSeek |
|---|---|
Parameters | 3.34 BVerifiedsafetensors.total from the HF API Source |
Licence | MITVerified |
Access | Public download |
Task (HF pipeline) | image-text-to-text |
First published on HF | 2025-10-17 |
Last updated on HF | 2025-11-04 |
Weights in memory | 6.7 GB at 16-bit · 2.0 GB at ~4.85 bitsEstimateParameters × bits ÷ 8. Excludes context cache and runtime overhead; not every model has a 4-bit build. |
Benchmark
75.7%
olmOCR-Bench, overall · #3 of 4 scored in this category
DeepSeek-OCR row (±1.0). Source
The description summarises the model card at https://huggingface.co/deepseek-ai/DeepSeek-OCR. Parameters and licence are read from the Hugging Face API; the weights-in-memory line is our arithmetic, not a measurement.