unlimited-ocr-gguf

Unlimited OCR GGUF

Unlimited OCR GGUF is the Edgework/Aether card for the upstream sahilchachra/Unlimited-OCR-GGUF document OCR release. It is admitted as an OCR lane, not a general chat model: requests must provide a page image plus the DeepSeek-OCR chat template prompt.

Artifacts

The Edgework qspec pins both files because the projector is required for image tokens. The default qspec prompt is:

<|grounding|>Convert the document to markdown.

llama.cpp

hf download sahilchachra/Unlimited-OCR-GGUF Unlimited-OCR-Q4_K_M.gguf mmproj-Unlimited-OCR-F16.gguf --local-dir ./uocr
llama-mtmd-cli \
  -m ./uocr/Unlimited-OCR-Q4_K_M.gguf \
  --mmproj ./uocr/mmproj-Unlimited-OCR-F16.gguf \
  --chat-template deepseek-ocr \
  --image ./page.png \
  --temp 0 \
  -n 512 \
  -p "<|grounding|>Convert the document to markdown." \
  --ignore-eos

Edgework Admission

This card is tied to open-source/gnosis/monster-swarm/qspecs/unlimited-ocr.qspec.json. The qspec renders a known PDF page to PNG, runs the GGUF + projector through llama-mtmd-cli, and gates on visible OCR text from the page. That proof is separate from chat-completion admission; unlimited-ocr remains outside the generic chat model list.

Cloudflare Mesh

The Cloudflare Worker mesh uses distributed-unlimited-ocr-coordinator as the coordinator service and two layer-node services. Worker config carries:

• MODEL_REPO_ID=sahilchachra/Unlimited-OCR-GGUF
• MODEL_GGUF=Unlimited-OCR-Q4_K_M.gguf
• MMPROJ_GGUF=mmproj-Unlimited-OCR-F16.gguf
• OCR_CHAT_TEMPLATE=deepseek-ocr

No Workers AI binding, external paid AI API, or always-on Cloud Run setting is part of this lane.