All tasks / Models
DeepSeek-OCR 2
DeepSeek's open-weights ~3B vision-language OCR model that converts document images to Markdown via compressed visual tokens.
- DeepSeek
- Open weights
- Released 27 Jan 2026
Where it ranks
- ADocuments & OCRModels · Local · 90 on OmniDocBench v1.6 (end-to-end, v1.6_full)
Facts
- Vendor
- DeepSeek
- Family
- DeepSeek-OCR
- Released
- 27 Jan 2026
- Weights
- Open: download and run it yourself
- License
- Apache-2.0
- Parameters
- 3.39B
- GPU memory to run locally
- ~9 GB (Estimate, no official figure: BF16 weights ≈ 3.39 B × 2 GB + ~1.5 GB runtime overhead.)
- Access
- Free (open weights)
- Last verified
- 18 Oct 2026
Price from huggingface.co, 18 Oct 2026.
Benchmark results
| Benchmark | Result | Entry on the leaderboard | As of | Source |
|---|---|---|---|---|
| OmniDocBench v1.6 (end-to-end, v1.6_full) | 90 | DeepSeek-OCR 2Overall score in the README's end-to-end table 'OmniDocBench (v1.6_full)' (last leaderboard update 2026-09-11). | 11 Sep 2026 | github.com |
| ParseBench (models) | 41.2% | DeepSeek-OCR-2Overall (mean of tables, charts, content faithfulness, semantic formatting, visual grounding) from leaderboard.csv in the ParseBench repo (last leaderboard commit 2026-10-09). | 9 Oct 2026 | github.com |
Where you can use it
No tracked tool lists this model yet; as open weights it runs in local runtimes.
Earlier versions
- DeepSeek-OCR · Oct 2025