---
language:
- th
base_model:
- Qwen/Qwen3.5-2B
license: apache-2.0
---

# Pathumma-llm-vision-3.0.0-preview 2B

**Pathumma-llm-vision-3.0.0-preview 2B** is a vision-language model developed by **NECTEC**, based on **Qwen3.5-2B** and further trained for Thai and multilingual OCR and image-text understanding.

The model was trained on **377K OCR samples** with a focus on improving OCR performance, particularly for Thai and challenging real-world document and scene-text images.


## Model Highlights

* 🧠 Based on **Qwen3.5-2B**
* 🇹🇭 Optimized for **Thai OCR**
* 📚 Trained on **377K OCR samples**
* 🖼️ Vision-language image-to-text understanding
* 📄 Designed for OCR and document understanding
* 🚀 Intended for efficient and compact deployment

## 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

| Model                                 | Document Parsing | Fine-grained Text Recognition | Full-page OCR | Handwritten Content Extraction | Text Recognition | Document Classification | Diagram VQA | Cognition VQA | Infographics |
| ------------------------------------- | ---------------: | ----------------------------: | ------------: | -----------------------------: | ---------------: | ----------------------: | ----------: | ------------: | -----------: |
| **Typhoon-OCR1.5-2B**                 |           0.2355 |                        0.1274 |    **0.7724** |                         0.2808 |           0.5922 |                  0.4093 |      0.4063 |        0.5690 |       0.5839 |
| **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 |
| **Pathumma-LLM-Vision-3.0.0-preview** |       **0.4872** |                        0.1458 |    **0.7831** |                         0.3286 |           0.6443 |                  0.2558 |  **0.5784** |    **0.6763** |   **0.6451** |

## Training

The model was trained using **377K OCR samples**.

### Training Configuration

| Parameter             | Value                             |
| --------------------- | --------------------------------- |
| Base model            | Qwen3.5-2B                        |
| Training data         | 377K OCR samples                  |
| Training method       | SFT                               |
| Learning rate         | `1.0e-5`                          |
| Epochs                | `3`                               |
|lr_scheduler           | cosine                            |
| Gradient accumulation | `2`                               |
| Hardware              | 4 × NVIDIA A100                   |



The training setup was designed for large-scale OCR fine-tuning using 4 NVIDIA A100 GPUs.


## Intended Use

Pathumma Vision 3.5-2B is intended for:

* Thai OCR
* Scene text recognition
* Document text extraction
* Thai document understanding
* Efficient OCR deployment

## Quickstart

### Installation

```bash
pip install -U transformers
pip install torch torchvision
```

### Using 🤗 Transformers

```python
import torch
from transformers import AutoProcessor, Qwen3_5ForConditionalGeneration

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

model = Qwen3_5ForConditionalGeneration.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:

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

## Acknowledgements

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

This model is built upon the Qwen3.5 architecture and benefits from the work of the Qwen team.

## Citation

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

```bibtex
@misc{PathummaVision3,
  author = {
    Phromchai, Theerawat and
    Kerdthaisong, Kun and
    Pintobtang, Thanaporn and
    Prachumkhong, Khemjira and
    Lilek, Teepakorn and
    Issaranon, Theerasit and
    Kongyoung, Sarawoot
  },
  title = {Pathumma Vision 3.5-2B},
  year = {2026},
  url = {https://huggingface.co/nectec/Pathumma-llm-vision-3.5-2b}
}

```