chinda-qwen3-4b/README.md

---
license: apache-2.0
language:
- th
- en
base_model:
- Qwen/Qwen3-4B
pipeline_tag: text-generation
tags:
- thai
---

# 🇹🇭 Chinda Opensource Thai LLM 4B

**Latest Model, Think in Thai, Answer in Thai, Built by Thai Startup**

![image/jpeg](https://cdn-uploads.huggingface.co/production/uploads/5fcd9c426d942eaf4d1ebd30/RTzTckBAT3MjYp950UamV.jpeg)

Chinda Opensource Thai LLM 4B is iApp Technology's cutting-edge Thai language model that brings advanced thinking capabilities to the Thai AI ecosystem. Built on the latest Qwen3-4B architecture, Chinda represents our commitment to developing sovereign AI solutions for Thailand.

## 🚀 Quick Links

- **🌐 Demo:** [https://chindax.iapp.co.th](https://chindax.iapp.co.th) (Choose ChindaLLM 4b)
- **📦 Model Download:** [https://huggingface.co/iapp/chinda-qwen3-4b](https://huggingface.co/iapp/chinda-qwen3-4b)
- **🐋 Ollama:** [https://ollama.com/iapp/chinda-qwen3-4b](https://ollama.com/iapp/chinda-qwen3-4b)
- **🏠 Homepage:** [https://iapp.co.th/products/chinda-opensource-llm](https://iapp.co.th/products/chinda-opensource-llm)
- **📄 License:** Apache 2.0

## ✨ Key Features

### 🆓 **0. Free and Opensource for Everyone**
Chinda LLM 4B is completely free and open-source, enabling developers, researchers, and businesses to build Thai AI applications without restrictions.

### 🧠 **1. Advanced Thinking Model**
- **Highest score among Thai LLMs in 4B category**
- Seamless switching between thinking and non-thinking modes
- Superior reasoning capabilities for complex problems
- Can be turned off for efficient general-purpose dialogue

### 🇹🇭 **2. Exceptional Thai Language Accuracy**
- **98.4% accuracy** in outputting Thai language
- Prevents unwanted Chinese and foreign language outputs
- Specifically fine-tuned for Thai linguistic patterns

### 🆕 **3. Latest Architecture**
- Based on the cutting-edge **Qwen3-4B** model
- Incorporates the latest advancements in language modeling
- Optimized for both performance and efficiency

### 📜 **4. Apache 2.0 License**
- Commercial use permitted
- Modification and distribution allowed
- No restrictions on private use

## 📊 Benchmark Results

Chinda LLM 4B demonstrates superior performance compared to other Thai language models in its category:

| Benchmark | Language | Chinda LLM 4B | Alternative* | 
|-----------|----------|---------------|-------------|
| **AIME24** | English | **0.533** | 0.100 |
| | Thai | **0.100** | 0.000 |
| **LiveCodeBench** | English | **0.665** | 0.209 |
| | Thai | **0.198** | 0.144 |
| **MATH500** | English | **0.908** | 0.702 |
| | Thai | **0.612** | 0.566 |
| **IFEVAL** | English | **0.849** | 0.848 |
| | Thai | 0.683 | **0.740** |
| **Language Accuracy** | Thai | 0.984 | **0.992** |
| **OpenThaiEval** | Thai | **0.651** | 0.544 |
| **AVERAGE** | | **0.569** | 0.414 |

* Alternative: scb10x_typhoon2.1-gemma3-4b
*  Tested by Skythought and Evalscope Benchmark Libraries by iApp Technology team. Results show Chinda LLM 4B achieving **37% better overall performance** than the nearest alternative.

## ✅ Suitable For

### 🔍 **1. RAG Applications (Sovereign AI)**
Perfect for building Retrieval-Augmented Generation systems that keep data processing within Thai sovereignty.

### 📱 **2. Mobile and Laptop Applications**
Reliable Small Language Model optimized for edge computing and personal devices.

### 🧮 **3. Math Calculation**
Excellent performance in mathematical reasoning and problem-solving.

### 💻 **4. Code Assistant**
Strong capabilities in code generation and programming assistance.

### ⚡ **5. Resource Efficiency**
Very fast inference with minimal GPU memory consumption, ideal for production deployments.

## ❌ Not Suitable For

### 📚 **Factual Questions Without Context**
As a 4B parameter model, it may hallucinate when asked for specific facts without provided context. Always use with RAG or provide relevant context for factual queries.

## 🛠️ Quick Start

### Installation

```bash
pip install transformers torch
```

### Basic Usage

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "iapp/chinda-qwen3-4b"

# Load the tokenizer and model
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto"
)

# Prepare the model input
prompt = "อธิบายเกี่ยวกับปัญญาประดิษฐ์ให้ฟังหน่อย"
messages = [
    {"role": "user", "content": prompt}
]

text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=True  # Enable thinking mode for better reasoning
)

model_inputs = tokenizer([text], return_tensors="pt").to(model.device)

# Generate response
generated_ids = model.generate(
    **model_inputs,
    max_new_tokens=1024,
    temperature=0.6,
    top_p=0.95,
    top_k=20,
    do_sample=True
)

output_ids = generated_ids[0][len(model_inputs.input_ids[0]):].tolist()

# Parse thinking content (if enabled)
try:
    # Find </think> token (151668)
    index = len(output_ids) - output_ids[::-1].index(151668)
except ValueError:
    index = 0

thinking_content = tokenizer.decode(output_ids[:index], skip_special_tokens=True).strip("\n")
content = tokenizer.decode(output_ids[index:], skip_special_tokens=True).strip("\n")

print("🧠 Thinking:", thinking_content)
print("💬 Response:", content)
```

### Switching Between Thinking and Non-Thinking Mode

#### Enable Thinking Mode (Default)
```python
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=True  # Enable detailed reasoning
)
```

#### Disable Thinking Mode (For Efficiency)
```python
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=False  # Fast response mode
)
```

### API Deployment

#### Using vLLM
```bash
pip install vllm>=0.8.5
vllm serve iapp/chinda-qwen3-4b --enable-reasoning --reasoning-parser deepseek_r1
```

#### Using SGLang
```bash
pip install sglang>=0.4.6.post1
python -m sglang.launch_server --model-path iapp/chinda-qwen3-4b --reasoning-parser qwen3
```

#### Using Ollama (Easy Local Setup)

**Installation:**
```bash
# Install Ollama (if not already installed)
curl -fsSL https://ollama.com/install.sh | sh

# Pull Chinda LLM 4B model
ollama pull iapp/chinda-qwen3-4b
```

**Basic Usage:**
```bash
# Start chatting with Chinda LLM
ollama run iapp/chinda-qwen3-4b

# Example conversation
ollama run iapp/chinda-qwen3-4b "อธิบายเกี่ยวกับปัญญาประดิษฐ์ให้ฟังหน่อย"
```

**API Server:**
```bash
# Start Ollama API server
ollama serve

# Use with curl
curl http://localhost:11434/api/generate -d '{
  "model": "iapp/chinda-qwen3-4b",
  "prompt": "สวัสดีครับ",
  "stream": false
}'
```

**Model Specifications:**
- **Size:** 2.5GB (quantized)
- **Context Window:** 40K tokens  
- **Architecture:** Optimized for local deployment
- **Performance:** Fast inference on consumer hardware


## 🔧 Advanced Configuration

### Processing Long Texts

Chinda LLM 4B natively supports up to 32,768 tokens. For longer contexts, enable YaRN scaling:

```json
{
    "rope_scaling": {
        "rope_type": "yarn",
        "factor": 4.0,
        "original_max_position_embeddings": 32768
    }
}
```

### Recommended Parameters

**For Thinking Mode:**
- Temperature: 0.6
- Top-P: 0.95
- Top-K: 20
- Min-P: 0

**For Non-Thinking Mode:**
- Temperature: 0.7
- Top-P: 0.8
- Top-K: 20
- Min-P: 0

## 📝 Context Length & Template Format

### Context Length Support
- **Native Context Length:** 32,768 tokens
- **Extended Context Length:** Up to 131,072 tokens (with YaRN scaling)
- **Input + Output:** Total conversation length supported
- **Recommended Usage:** Keep conversations under 32K tokens for optimal performance

### Chat Template Format

Chinda LLM 4B uses a standardized chat template format for consistent interactions:

```python
# Basic template structure
messages = [
    {"role": "system", "content": "You are a helpful Thai AI assistant."},
    {"role": "user", "content": "สวัสดีครับ"},
    {"role": "assistant", "content": "สวัสดีค่ะ! มีอะไรให้ช่วยเหลือบ้างคะ"},
    {"role": "user", "content": "ช่วยอธิบายเรื่อง AI ให้ฟังหน่อย"}
]

# Apply template with thinking mode
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=True
)
```

### Template Structure

The template follows the standard conversational format:

```
<|im_start|>system
You are a helpful Thai AI assistant.<|im_end|>
<|im_start|>user
สวัสดีครับ<|im_end|>
<|im_start|>assistant
สวัสดีค่ะ! มีอะไรให้ช่วยเหลือบ้างคะ<|im_end|>
<|im_start|>user
ช่วยอธิบายเรื่อง AI ให้ฟังหน่อย<|im_end|>
<|im_start|>assistant
```

### Advanced Template Usage

```python
# Multi-turn conversation with thinking control
def create_conversation(messages, enable_thinking=True):
    # Add system message if not present
    if not messages or messages[0]["role"] != "system":
        system_msg = {
            "role": "system", 
            "content": "คุณเป็น AI ผู้ช่วยที่ฉลาดและเป็นประโยชน์ พูดภาษาไทยได้อย่างเป็นธรรมชาติ"
        }
        messages = [system_msg] + messages
    
    # Apply chat template
    text = tokenizer.apply_chat_template(
        messages,
        tokenize=False,
        add_generation_prompt=True,
        enable_thinking=enable_thinking
    )
    
    return text

# Example usage
conversation = [
    {"role": "user", "content": "คำนวณ 15 × 23 = ?"},
]

prompt = create_conversation(conversation, enable_thinking=True)
```

### Dynamic Mode Switching

You can control thinking mode within conversations using special commands:

```python
# Enable thinking for complex problems
messages = [
    {"role": "user", "content": "/think แก้สมการ: x² + 5x - 14 = 0"}
]

# Disable thinking for quick responses  
messages = [
    {"role": "user", "content": "/no_think สวัสดี"}
]
```

### Context Management Best Practices

1. **Monitor Token Count:** Keep track of total tokens (input + output)
2. **Truncate Old Messages:** Remove oldest messages when approaching limits
3. **Use YaRN for Long Contexts:** Enable rope scaling for documents > 32K tokens
4. **Batch Processing:** For very long texts, consider chunking and processing in batches

```python
def manage_context(messages, max_tokens=30000):
    """Simple context management function"""
    total_tokens = sum(len(tokenizer.encode(msg["content"])) for msg in messages)
    
    while total_tokens > max_tokens and len(messages) > 2:
        # Keep system message and remove oldest user/assistant pair
        if messages[1]["role"] == "user":
            messages.pop(1)  # Remove user message
            if len(messages) > 1 and messages[1]["role"] == "assistant":
                messages.pop(1)  # Remove corresponding assistant message
        
        total_tokens = sum(len(tokenizer.encode(msg["content"])) for msg in messages)
    
    return messages
```

## 🏢 Enterprise Support

For enterprise deployments, custom training, or commercial support, contact us at:
- **Email:** sale@iapp.co.th
- **Website:** [iapp.co.th](https://iapp.co.th)

## ❓ Frequently Asked Questions

<details>
<summary><strong>🏷️ Why is it named "Chinda"?</strong></summary>

The name "Chinda" (จินดา) comes from "จินดามณี" (Chindamani), which is considered the first book of Thailand written by Phra Horathibodi (Sri Dharmasokaraja) in the Sukhothai period. Just as จินดามณี was a foundational text for Thai literature and learning, Chinda LLM represents our foundation for Thai sovereign AI - a model that truly understands and thinks in Thai, preserving and advancing Thai language capabilities in the digital age.

</details>

<details>
<summary><strong>⚖️ Can I use Chinda LLM 4B for commercial purposes?</strong></summary>

Yes! Chinda LLM 4B is released under the **Apache 2.0 License**, which allows:
- ✅ **Commercial use** - Use in commercial products and services
- ✅ **Research use** - Academic and research applications
- ✅ **Modification** - Adapt and modify the model
- ✅ **Distribution** - Share and redistribute the model
- ✅ **Private use** - Use for internal company projects

No restrictions on commercial applications - build and deploy freely!

</details>

<details>
<summary><strong>🧠 What's the difference between thinking and non-thinking mode?</strong></summary>

**Thinking Mode (`enable_thinking=True`):**
- Model shows its reasoning process in `<think>...</think>` blocks
- Better for complex problems, math, coding, logical reasoning
- Slower but more accurate responses
- Recommended for tasks requiring deep analysis

**Non-Thinking Mode (`enable_thinking=False`):**
- Direct answers without showing reasoning
- Faster responses for general conversations
- Better for simple queries and chat applications
- More efficient resource usage

You can switch between modes or let users control it dynamically using `/think` and `/no_think` commands.

</details>

<details>
<summary><strong>📊 How does Chinda LLM 4B compare to other Thai language models?</strong></summary>

Chinda LLM 4B achieves **37% better overall performance** compared to the nearest alternative:

- **Overall Average:** 0.569 vs 0.414 (alternative)
- **Math (MATH500):** 0.908 vs 0.702 (English), 0.612 vs 0.566 (Thai)
- **Code (LiveCodeBench):** 0.665 vs 0.209 (English), 0.198 vs 0.144 (Thai)
- **Thai Language Accuracy:** 98.4% (prevents Chinese/foreign text output)
- **OpenThaiEval:** 0.651 vs 0.544

It's currently the **highest-scoring Thai LLM in the 4B parameter category**.

</details>

<details>
<summary><strong>💻 What are the system requirements to run Chinda LLM 4B?</strong></summary>

**Minimum Requirements:**
- **GPU:** 8GB VRAM (RTX 3070/4060 Ti or better)
- **RAM:** 16GB system memory
- **Storage:** 8GB free space for model download
- **Python:** 3.8+ with PyTorch

**Recommended for Production:**
- **GPU:** 16GB+ VRAM (RTX 4080/A4000 or better)
- **RAM:** 32GB+ system memory
- **Storage:** SSD for faster loading

**CPU-Only Mode:** Possible but significantly slower (not recommended for production)

</details>

<details>
<summary><strong>🔧 Can I fine-tune Chinda LLM 4B for my specific use case?</strong></summary>

Yes! As an open-source model under Apache 2.0 license, you can:

- **Fine-tune** on your domain-specific data
- **Customize** for specific tasks or industries
- **Modify** the architecture if needed
- **Create derivatives** for specialized applications

Popular fine-tuning frameworks that work with Chinda:
- **Unsloth** - Fast and memory-efficient
- **LoRA/QLoRA** - Parameter-efficient fine-tuning
- **Hugging Face Transformers** - Full fine-tuning
- **Axolotl** - Advanced training configurations

Need help with fine-tuning? Contact our team at sale@iapp.co.th

</details>

<details>
<summary><strong>🌍 What languages does Chinda LLM 4B support?</strong></summary>

**Primary Languages:**
- **Thai** - Native-level understanding and generation (98.4% accuracy)
- **English** - Strong performance across all benchmarks

**Additional Languages:**
- 100+ languages supported (inherited from Qwen3-4B base)
- Focus optimized for Thai-English bilingual tasks
- Code generation in multiple programming languages

**Special Features:**
- **Code-switching** between Thai and English
- **Translation** between Thai and other languages
- **Multilingual reasoning** capabilities

</details>

<details>
<summary><strong>🔍 Is the training data publicly available?</strong></summary>

The model weights are open-source, but the specific training datasets are not publicly released. However:

- **Base Model:** Built on Qwen3-4B (Alibaba's open foundation)
- **Thai Optimization:** Custom dataset curation for Thai language tasks
- **Quality Focus:** Carefully selected high-quality Thai content
- **Privacy Compliant:** No personal or sensitive data included

For research collaborations or dataset inquiries, contact our research team.

</details>

<details>
<summary><strong>🆘 How do I get support or report issues?</strong></summary>

**For Technical Issues:**
- **GitHub Issues:** Report bugs and technical problems
- **Hugging Face:** Model-specific questions and discussions
- **Documentation:** Check our comprehensive guides

**For Commercial Support:**
- **Email:** sale@iapp.co.th
- **Enterprise Support:** Custom training, deployment assistance
- **Consulting:** Integration and optimization services

**Community Support:**
- **Thai AI Community:** Join discussions about Thai AI development
- **Developer Forums:** Connect with other Chinda users

</details>

<details>
<summary><strong>📥 How large is the model download and what format is it in?</strong></summary>

**Model Specifications:**
- **Parameters:** 4.02 billion (4B)
- **Download Size:** ~8GB (compressed)
- **Format:** Safetensors (recommended) and PyTorch
- **Precision:** BF16 (Brain Float 16)

**Download Options:**
- **Hugging Face Hub:** `huggingface.co/iapp/chinda-qwen3-4b`
- **Git LFS:** For version control integration
- **Direct Download:** Individual model files
- **Quantized Versions:** Available for reduced memory usage (GGUF, AWQ)

**Quantization Options:**
- **4-bit (GGUF):** ~2.5GB, runs on 4GB VRAM
- **8-bit:** ~4GB, balanced performance/memory
- **16-bit (Original):** ~8GB, full performance

</details>

## 📚 Citation

If you use Chinda LLM 4B in your research or projects, please cite:

```bibtex
@misc{chinda-llm-4b,
  title={Chinda LLM 4B: Thai Sovereign AI Language Model},
  author={iApp Technology},
  year={2025},
  publisher={Hugging Face},
  url={https://huggingface.co/iapp/chinda-qwen3-4b}
}
```

---

*Built with 🇹🇭 by iApp Technology - Empowering Thai Businesses with Sovereign AI Excellence*

![image/jpeg](https://cdn-uploads.huggingface.co/production/uploads/5fcd9c426d942eaf4d1ebd30/qNa4bznh179myghTFcpFp.jpeg)

**Powered by iApp Technology**

<i>Disclaimer: Provided responses are not guaranteed.</i>