ChindaLLM 4B

What it does​

• Thinks before it answersA reasoning mode for maths and code, switched off for a faster direct reply.
• Answers in ThaiA Thai question gets its answer in Thai, not in English.
• Runs on your machineThe 2.5 GB quantized build runs on a laptop; 8 GB of memory is recommended.
• Reads your documentsGive it the passage and it answers from it, with no data leaving your network.

Measured​

Against the leading alternative Thai model of the same size. Higher is better.

In the full comparison on the Hugging Face model card, Chinda 4B trails only on Thai IFEval, above, and on Thai language accuracy, 0.984 against 0.992.

Run it​

ollama run iapp/chinda-qwen3-4b

from transformers import AutoModelForCausalLM, AutoTokenizer

name = "iapp/chinda-qwen3-4b"
tok = AutoTokenizer.from_pretrained(name)
model = AutoModelForCausalLM.from_pretrained(name, torch_dtype="auto", device_map="auto")

messages = [{"role": "user", "content": "สรุปเอกสารนี้ให้หน่อย"}]
text = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True, enable_thinking=True)
out = model.generate(**tok([text], return_tensors="pt").to(model.device), max_new_tokens=2048)
print(tok.decode(out[0], skip_special_tokens=True))

Step-by-step guides: LM Studio · Open WebUI · Ollama · n8n

Limits​

• Not for factual questions asked with no context. A four-billion-parameter model is not an encyclopaedia and can state facts wrongly; give it the document and it reasons over it well.
• The reasoning pass takes time. Switch it off when a plain reply will do.

Licence​

Apache 2.0: free to use, modify and deploy, including commercially, with no usage reporting.