加载中…
加载中…
GGUF conversion of lightonai/LightOnOCR-3-0.8B (Qwen3.5 vision-language architecture, Apache 2.0). All credit for the model goes to LightOn; this repo only contains format conversions.
| File | Description | Size |
|---|---|---|
LightOnOCR-3-0.8B-F16.gguf |
text model, F16 | 1.5 GB |
LightOnOCR-3-0.8B-Q8_0.gguf |
text model, Q8_0 | 0.8 GB |
mmproj-LightOnOCR-3-0.8B-F16.gguf |
vision projector (required for images) | 0.2 GB |
Converted with llama.cpp convert_hf_to_gguf.py (--outtype f16 --no-mtp, mmproj with --mmproj),
Q8_0 made with llama-quantize (build b12290).
Prompt modes (from the original card): empty prompt = plain transcription, grounding = blocks with
labels and bounding boxes. Other instructions are out of distribution.
llama-mtmd-cli -m LightOnOCR-3-0.8B-Q8_0.gguf --mmproj mmproj-LightOnOCR-3-0.8B-F16.gguf --image page.png -p ""
Download the files and the Modelfile into one folder, then:
ollama create lightonocr-3:0.8b -f Modelfile
Disable thinking when calling the model ("think": false in the API, --think=false in the CLI).
The Modelfile uses temperature 0 and the built-in qwen3.5 renderer/parser.
Tested only on synthetic invoice images (Polish text); no benchmarks were run on the conversions.