LightOnOCR-3-4B-MTP-Repacked

LightOnOCR-3-4B, developed by lightonai, is the largest and most accurate model in the LightOnOCR-3 family of end-to-end OCR vision-language models, built on the Qwen3.5-4B architecture and released under the Apache 2.0 license. Recommended for most OCR tasks, it serves as a powerful, single-model alternative to complex document understanding pipelines by combining high-quality text transcription (triggered by an empty prompt) with advanced visual understanding features. When used with the grounding prompt, the model outputs labeled bounding boxes for all document elements, generates short descriptions for images, and extracts numerical data from charts into structured HTML tables. Highly versatile and optimized for seamless integration with Transformers and vLLM, it excels at processing complex layouts—including tables, receipts, forms, multi-column documents, and math notation—and performs best when documents are preprocessed at 400 DPI with a target longest dimension of 2048px.

MTP — Speculative Decoding: Yes | GGUF — LightOnOCR-3-4B-MTP-Repacked(GGUF)

Multi-Token Prediction (MTP) GGUF is a specialized GGUF model file format extension that integrates speculative decoding directly into the model weights to significantly accelerate local inference. Unlike traditional speculative decoding which requires a separate, smaller "draft" model, MTP GGUF files include additional output heads within the main model architecture that predict multiple future tokens in a single forward pass.

Model Files

File Name	Quant Type	File Size	File Link	Description
LightOnOCR-3-4B-MTP.BF16.gguf	BF16	8.67 GB	Link	Full BF16 weights. Highest quality, largest file size.
LightOnOCR-3-4B-MTP.Q3_K_L.gguf	Q3_K_L	2.48 GB	Link	Lower quality but usable, good for low RAM availability.
LightOnOCR-3-4B-MTP.Q3_K_M.gguf	Q3_K_M	2.32 GB	Link	Low quality.
LightOnOCR-3-4B-MTP.Q4_K_M.gguf	Q4_K_M	2.78 GB	Link	Good quality, default size for most use cases, recommended.
LightOnOCR-3-4B-MTP.Q4_K_S.gguf	Q4_K_S	2.63 GB	Link	Slightly lower quality with more space savings, recommended.
LightOnOCR-3-4B-MTP.Q5_K_M.gguf	Q5_K_M	3.16 GB	Link	High quality, recommended.
LightOnOCR-3-4B-MTP.Q5_K_S.gguf	Q5_K_S	3.07 GB	Link	High quality, recommended.
LightOnOCR-3-4B-MTP.Q6_K.gguf	Q6_K	3.56 GB	Link	Very high quality, near perfect, recommended.
LightOnOCR-3-4B-MTP.Q8_0.gguf	Q8_0	4.61 GB	Link	Extremely high quality, generally unneeded but max available quant.
LightOnOCR-3-4B-MTP.mmproj-bf16.gguf	mmproj-bf16	676 MB	Link	Multimodal projection file in BF16 format. Used for vision/OCR capabilities.

MTP-Repacked [Transformers, BF16]

File Name	File Type	File Size	File Link	Description
chat_template.jinja	Template	7.76 kB	Link	Chat template defining the model's conversation and prompting format.
config.json	Config	2.9 kB	Link	Main model architecture and hyperparameter configuration.
generation_config.json	Config	204 Bytes	Link	Default parameters controlling text generation (e.g., temperature, top_p, max_length).
model-00001-of-00003.safetensors	Weights	4.35 GB	Link	Model weight shard 1/3. Required to load the model.
model-00002-of-00003.safetensors	Weights	4.39 GB	Link	Model weight shard 2/3. Required to load the model.
model-00003-of-00003.safetensors	Weights	572 MB	Link	Model weight shard 3/3. Required to load the model.
model.safetensors.index.json	Index	67.3 kB	Link	Index file mapping model weights to their respective shard files.
preprocessor_config.json	Config	390 Bytes	Link	Configuration for input preprocessing (e.g., image resizing, normalization).
processor_config.json	Config	1.19 kB	Link	Configuration for the multimodal processor combining vision and text inputs.
tokenizer.json	Tokenizer	20 MB	Link	Fast tokenizer vocabulary, merge rules, and special tokens.
tokenizer_config.json	Tokenizer	1.14 kB	Link	Tokenizer initialization parameters and settings.
vocab.json	Tokenizer	6.72 MB	Link	Vocabulary mapping (token IDs to strings) for the tokenizer.

BF16 Test Run

Tested LightOnOCR-3-4B-MTP.BF16.gguf (with LightOnOCR-3-4B-MTP.mmproj-bf16.gguf) on an NVIDIA RTX PRO 6000 Blackwell Server Edition GPU with 96 GB of VRAM, using llama-cli with full GPU offloading and MTP speculative decoding.

Performance Metrics

Metric	Value
Average Prompt Processing Speed	2,453.59 tokens/s
Average Generation Speed	208.55 tokens/s
Total Wall Time	23.2 seconds
Context Length	8,192 tokens
KV Cache (K)	q8_0
GPU	NVIDIA RTX PRO 6000 Blackwell Server Edition
GPU VRAM Capacity	96 GB
Peak GPU VRAM Usage	10.42 GiB
Peak GPU Utilization	88%
GPU Layers Offloaded	99 (-ngl 99)
Random Seed	42
Speculative Decoding	MTP (draft_n_max=2)
Test Runs	3

Test Shots

Shot	Test
Shot 1	Vision OCR — Image 1
Shot 2	Vision OCR — Image 2
Shot 3	Vision OCR — Image 3

Run Inference

./build/bin/llama-cli \
  -m /path/to/models/LightOnOCR-3-4B-MTP.BF16.gguf \
  --mmproj /path/to/models/LightOnOCR-3-4B-MTP.mmproj-bf16.gguf \
  --image /path/to/image.jpeg \
  --spec-type draft-mtp \
  --spec-draft-n-max 2 \
  -ngl 99 \
  -c 8192 \
  --cache-type-k q8_0 \
  --seed 42 \
  -n 1024 \
  -st \
  -p "OCR the image."

Performance results reflect this specific hardware and inference configuration. Actual throughput may vary depending on the prompt, model, software build, and runtime settings.

Snapshot(3)

llama.cpp

LLM inference in C/C++ — https://github.com/ggml-org/llama.cpp

MTP-repackaged and quantized using llama.cpp by — prithivMLmods.