nectec

Pathumma-llm-vision-3.0.0-re

README

License: apache-2.0

Model Highlights

🧠 Based on Qwen3-VL-2B-Instruct
🇹🇭 Optimized for Thai OCR
📚 Trained on 377K OCR samples
⚡ Quantization-Aware Training (QAT)
🖼️ Vision-language image-to-text understanding
📄 Designed for OCR and document/image understanding
🚀 Intended for efficient deployment and inference

Benchmark

We evaluate our models on ThaiOCRBench, a benchmark designed to assess OCR and document understanding capabilities across Thai and challenging real-world visual content.

ThaiOCRBench Results

Table with columns: Model, Document Parsing, Fine-grained Text Recognition, Full-page OCR, Handwritten Content Extraction, Text Recognition, Document Classification, Diagram VQA, Cognition VQA, Infographics, Overall
Model	Document Parsing	Fine-grained Text Recognition	Full-page OCR	Handwritten Content Extraction	Text Recognition	Document Classification	Diagram VQA	Cognition VQA	Infographics	Overall
Typhoon-OCR1.5-2B	0.2355	0.1274	0.7724	0.2808	0.5922	0.4093	0.4063	0.5690	0.5839	0.4419
Pathumma-LLM-Vision-3.0.0-re	0.4117	0.1478	0.7361	0.3675	0.6742	0.4326	0.4363	0.6623	0.5951	0.4960

Training

The model was fine-tuned on 377K OCR training samples.

Training Configuration

Table with columns: Parameter, Value
Parameter	Value
Base model	Qwen3-VL-2B-Instruct
Training data	377K OCR samples
Training method	Quantization-Aware Training (QAT)
Learning rate	`9e-6`
Epochs	`2`
Gradient accumulation	`4`
Hardware	8 × NVIDIA A100

Training Hardware

Training was performed using:

8 × NVIDIA A100 GPUs

with gradient accumulation of 4.

Intended Use

Pathumma Vision 3.0.0-Re is intended for:

Thai OCR
Scene text recognition
Document text extraction
Thai document understanding

Quickstart

Installation

bash
pip install -U transformers
pip install torch torchvision

Using 🤗 Transformers

python
import torch
from transformers import Qwen3VLForConditionalGeneration, AutoProcessor

model_id = "nectec/Pathumma-llm-vision-3.0.0-re"

model = Qwen3VLForConditionalGeneration.from_pretrained(
    model_id,
    dtype=torch.bfloat16,
    device_map="auto",
)

processor = AutoProcessor.from_pretrained(model_id)

messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "image",
                "image": "path/to/your/image.jpg",
            },
            {
                "type": "text",
                "text": "อ่านข้อความในภาพนี้",
            },
        ],
    }
]

inputs = processor.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    return_dict=True,
    return_tensors="pt",
)

inputs = inputs.to(model.device)

generated_ids = model.generate(
    **inputs,
    max_new_tokens=512,
)

generated_ids_trimmed = [
    out_ids[len(in_ids):]
    for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
]

output_text = processor.batch_decode(
    generated_ids_trimmed,
    skip_special_tokens=True,
    clean_up_tokenization_spaces=False,
)

print(output_text[0])

Contributors

This model was developed by:

Kun Kerdthaisong
Thanaporn Pintobtang
Theerawat Phromchai
Khemjira Prachumkhong
Teepakorn Lilek
Theerasit Issaranon
Sarawoot Kongyoung

Acknowledgements

We thank the NECTEC team and contributors involved in the development of Pathumma and the underlying Thai-language and vision-language resources.

This model is built upon the Qwen3-VL architecture and benefits from the work of the Qwen team.

Citation

If you find Pathumma-llm-vision-3.0.0-re useful in your research, please cite:

bibtex
@misc{PathummaVision3,
  author = {
    Kerdthaisong, Kun and
    Pintobtang, Thanaporn and
    Phromchai, Theerawat and
    Prachumkhong, Khemjira and
    Lilek, Teepakorn and
    Issaranon, Theerasit and
    Kongyoung, Sarawoot
  },
  title = {Pathumma Vision 3.0.0-Re},
  year = {2026},
  url = {https://huggingface.co/nectec/Pathumma-llm-vision-3.0.0-re}
}

Please also cite the original Qwen3-VL work:

bibtex
@misc{qwen3technicalreport,
  title={Qwen3 Technical Report},
  author={Qwen Team},
  year={2025},
  eprint={2505.09388},
  archivePrefix={arXiv},
  primaryClass={cs.CL},
  url={https://arxiv.org/abs/2505.09388}
}

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Container

Run this model inference with full control and performance in your environment.

Learn more

Model Details

Model Provider

nectec

Model Tree

Base

Qwen/Qwen3-VL-2B-Instruct

Fine-tuned

this model

Input Modalities

TextImage

Output Modalities

Text

Supported Functionality

Dedicated EndpointsContainer

Explore FriendliAI today

Get started Talk to an engineer