What it is#
Chinda LLM 4B is iApp Technology’s open-weight Thai sovereign large language model, developed in collaboration with AIEAT (the AI Engineering Association of Thailand). Built on the Qwen3-4B architecture and optimized for Thai, it brings strong Thai language understanding and generation, plus step-by-step thinking-mode reasoning for math, logic, and code, into an efficient 4-billion-parameter model you can run yourself.
The weights are open under Apache-2.0, so anyone can download, run, and fine-tune the model freely — for research or in production.
Key capabilities#
- Thai-first — trained and tuned specifically for Thai language understanding and generation.
- Thinking mode — explicit step-by-step reasoning for math, code, and logic problems.
- Long context — handles up to 40,960 tokens, enough for long Thai documents.
- Efficient — a 4B model runs on a single consumer GPU (~8 GB), or locally via Ollama / LM Studio / OpenWebUI.
- Open & OpenAI-compatible — open weights under Apache-2.0, and an OpenAI-compatible API for drop-in integration.
Benchmarks#
Evaluated head-to-head against another strong 4B Thai model, Typhoon 2.1 Gemma3 4B, Chinda LLM 4B leads on overall average and on most reasoning, code, and Thai-exam benchmarks.
| Benchmark | Language | Chinda 4B | Typhoon 2.1 4B |
|---|---|---|---|
| AIME24 | English | 0.533 | 0.100 |
| AIME24 | Thai | 0.100 | 0.000 |
| LiveCodeBench | English | 0.665 | 0.209 |
| LiveCodeBench | Thai | 0.198 | 0.144 |
| MATH500 | English | 0.908 | 0.702 |
| MATH500 | Thai | 0.612 | 0.566 |
| IFEval | English | 0.849 | 0.848 |
| IFEval | Thai | 0.683 | 0.740 |
| Language Accuracy (code-switching) | Thai | 0.984 | 0.992 |
| OpenThaiEval | Thai | 0.651 | 0.544 |
| Average | 0.569 | 0.414 |
Source: OpenThaiGPT benchmark suite (2024–2025). Higher is better.
How to use#
Load the model with the Hugging Face Transformers library and apply the chat template:
from transformers import AutoModelForCausalLM, AutoTokenizer
model_name = "iapp/chinda-qwen3-4b"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name, device_map="auto")
messages = [
{"role": "user", "content": "สวัสดีครับ ช่วยอธิบายเกี่ยวกับปัญญาประดิษฐ์ให้ฟังหน่อย"}
]
text = tokenizer.apply_chat_template(
messages, tokenize=False, add_generation_prompt=True
)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512, temperature=0.7)
print(tokenizer.decode(outputs[0][inputs.input_ids.shape[-1]:], skip_special_tokens=True))It also runs locally with Ollama, LM Studio, and OpenWebUI, and is available through an OpenAI-compatible API.
Use cases#
- Conversational AI — natural Thai chatbots and virtual assistants.
- Content generation — Thai articles, summaries, and marketing copy.
- Document understanding — read and summarize long Thai documents.
- Reasoning — math, logic, and code with step-by-step thinking.
- Education — Thai-language tutoring and explanations.
Availability#
Chinda LLM 4B on Hugging Face. Open weights, Apache-2.0.