LightOnOCR 3 0.8B MTP Stage needs about 1.5 GB of memory at Q4_K_M (a 0.5 GB file) with an 8K context window, so it fits on a 4 GB GPU. Longer context needs more: at 32K tokens the KV cache alone adds about 1.7 GB.
LightOnOCR 3 0.8B MTP Stage – lightweight OCR with vision and grounding
LightOnOCR 3 0.8B MTP Stage is a small OCR model based on the Qwen3.5 vision‑language architecture. It supports transcription, grounding with bounding boxes, and visual understanding of images, charts, and tables. The model is released under Apache 2.0 for research and commercial use.
Good at
- Fast transcription of documents
- Grounding with labels and bounding boxes
- Visual description of images and charts
Limits
- No vision support in older llama.cpp versions
- Requires MTP decoding for vision
- Limited to 262144 context tokens
Best for
- Transcribing receipts and forms
- Extracting tables from PDFs
- Generating image captions
Who it’s for. Developers building document‑understanding pipelines who need a single lightweight model for text, layout, and visual data extraction. The model is suitable for research and commercial applications under the Apache 2.0 license.
Quants & memory
| Quant | File | Est. memory | vs VRAM | Result | |
|---|---|---|---|---|---|
| Q4_K_M | 0.5 GB | 1.5 GB | Fits | Get file → |
The Q4_K_M quantization file is 0.54 GB and requires about 1.5 GB of VRAM at 8 k context. This is the only quant available, so it is the recommended choice for users with limited GPU memory.
What GPU do you need for LightOnOCR 3 0.8B MTP Stage?
These figures add the file size, an fp16 KV cache at an 8K context window (worked out from this model’s own header) and about 0.5 GB for the runtime. They are estimates, so leave a little headroom, and expect to use less context or a smaller quant if you are right at the limit.
- 4 GB VRAM: the best quant that fits is Q4_K_M (0.5 GB file, about 1.5 GB in use).
Which LightOnOCR 3 0.8B MTP Stage quant should you choose?
Q4_K_M is the usual starting point for LightOnOCR 3 0.8B MTP Stage: it balances file size and quality well. Here it is a 0.5 GB file.
New to quant names? Read GGUF quantization explained.
Specifications
Architecture from the GGUF header
Source
Frequently asked questions about LightOnOCR 3 0.8B MTP Stage
How much VRAM does LightOnOCR 3 0.8B MTP Stage need?
About 1.5 GB at Q4_K_M with an 8K context window. That is the 0.5 GB file, plus the KV cache and roughly 0.5 GB of runtime overhead. Smaller quants need less and longer context needs more.
Can I run LightOnOCR 3 0.8B MTP Stage on an 8 GB GPU?
Yes, with Q4_K_M: the 0.5 GB file needs about 1.5 GB at 8K context. Larger quants in this repository do not fit in 8 GB.
What is the file size of LightOnOCR 3 0.8B MTP Stage GGUF?
It depends on the quant: Q4_K_M is 0.5 GB.
What context length does LightOnOCR 3 0.8B MTP Stage support?
The GGUF header lists a trained context length of 262,144 tokens (256K). The KV cache grows with the context you use: about 0.4 GB at 8K and 1.7 GB at 32K.
Where can I download LightOnOCR 3 0.8B MTP Stage GGUF?
The files are on Hugging Face in the staghado/LightOnOCR-3-0.8B-GGUF repository. The download buttons on this page open the file on Hugging Face. ggufmodels.com does not host any model files.
What license does LightOnOCR 3 0.8B MTP Stage use?
Hugging Face lists the license as apache-2.0. Licenses can differ between the quantized files and the original model, so check the repository and the base model before commercial use.
Can I run LightOnOCR 3 0.8B on 8 GB VRAM?
Yes, the Q4_K_M quantization uses approximately 1.5 GB of VRAM at 8 k context, leaving ample headroom in an 8 GB GPU.
Does the model support vision in llama.cpp?
Vision support requires llama.cpp 0.6.0 or newer with MTP decoding enabled.
What license applies to this model?
The model is released under the Apache 2.0 license, allowing research and commercial use.