cstr/jina-ocr-v1-GGUF

jina-ocr-v1 — GGUF for CrispEmbed

GGUF conversion of jinaai/jina-ocr-v1 for CrispEmbed: a 3.4B mixture-of-experts OCR model (~570M active parameters). It uses DeepSeek-OCR's architecture: SAM ViT-B + CLIP-L/14 vision encoder and a DeepSeek-V2 MoE decoder.

Licence: CC BY-NC 4.0 — non-commercial use only. © Jina AI. For commercial use, contact Jina AI. These files are a format conversion of the original weights and carry the same licence and attribution requirements.

file	size
jina-ocr-v1-f16.gguf	6.7 GB
jina-ocr-v1-q8_0.gguf	3.6 GB
jina-ocr-v1-q4_k.gguf	2.3 GB

crispembed -m jina-ocr-v1 --ocr page.png

The GGUF stores the model's own prompt ("Transcribe the provided document image into a clean Markdown format, preserving the natural reading order.", in its <|User|> / <|Assistant|> chat layout), its rope base (1e6), and plain greedy decoding, so no extra flags are needed.

Parity

Checked against transformers (trust_remote_code, the model's own processor) with tools/ci-heavy/jina_ocr_v1.py, on the same 1024 × 1024 global view (upstream's crop_mode=False, base_size=1024). The text was identical on both test fixtures, a single line and a full book page, including when upstream ran in float32 on CrispEmbed's own preprocessed pixels.

CrispEmbed encodes one global view. Upstream's default "Gundam" mode also adds up to nine 640 × 640 tiles, which helps on dense, small print; that mode is not implemented here.