LibreDeepSeekOCR2

DeepSeek-OCR-2 (3B mixture-of-experts document OCR model with a Qwen2-based "visual causal flow" encoder) weights mirrored for LibreYOLO's ocr task.

Source

Mirrored from deepseek-ai/DeepSeek-OCR-2 at commit aaa02f3811945a91062062994c5c4a3f4c0af2b0 (code, paper arXiv:2601.20552). Licensed under the Apache License 2.0.

Modifications

No learned parameters were changed. The safetensors, config.json, tokenizer and processor files are byte-identical to the upstream snapshot, and the upstream LICENSE.txt is shipped verbatim as LICENSE. The upstream remote-code Python files and example assets are not included: LibreYOLO runs its own PyTorch port of the architecture and never executes remote code.

Usage

from libreyolo import LibreVLM

model = LibreVLM("deepseek-ocr-2", prompt="markdown")
page = model.predict("page.png")
print(page.ocr.markdown)

License

Apache License 2.0. See LICENSE and NOTICE.