The smallest OpenThai model and the easiest place to start — a 7B Thai chat model that runs on a laptop. Fine-tuned on over 2,000,000 Thai instruction pairs, with multi-turn conversation, RAG and tool-calling support.

If you want reasoning, use R1 32B. If you want Thai law, use OpenThai 2.0 Legal.

Quants

File	Quant	Size	Notes
`openthaigpt1.5-7b-instruct.Q4_K_M.gguf`	Q4_K_M	~4.7 GB	Recommended. Runs on 8 GB of RAM or a small GPU.
`openthaigpt1.5-7b-instruct.Q5_K_M.gguf`	Q5_K_M	~5.4 GB	Better quality, still laptop-friendly.
`openthaigpt1.5-7b-instruct.Q8_0.gguf`	Q8_0	~8.1 GB	Near-lossless.

Usage

Ollama

ollama run hf.co/openthaigpt/openthaigpt1.5-7b-instruct-GGUF:Q4_K_M

llama.cpp

bash

llama-cli -m openthaigpt1.5-7b-instruct.Q4_K_M.gguf \
  -p "ช่วยสรุปข้อดีข้อเสียของรถยนต์ไฟฟ้าในบริบทประเทศไทย" -n 2048 --temp 0.7

This model uses the ChatML template (<|im_start|> / <|im_end|>), embedded in the GGUF. Recommended sampling: temperature=0.7, top_p=0.9.

Capabilities

Multi-turn conversation
Retrieval-Augmented Generation — pair with OpenThaiRAG
Tool calling
Up to 131,072 input tokens in the full-precision model

Limitations

At 7B this is the least capable model in the family — it trades accuracy for the ability to run anywhere. For anything where correctness matters more than convenience, use a larger model. Not safety-tuned for open consumer deployment without your own guardrails.

Citation

bibtex

@misc{yuenyong2024openthaigpt15thaicentric,
      title={OpenThaiGPT 1.5: A Thai-Centric Open Source Large Language Model},
      author={Sumeth Yuenyong and Kobkrit Viriyayudhakorn and Apivadee Piyatumrong and Jillaphat Jaroenkantasima},
      year={2024},
      eprint={2411.07238},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2411.07238}
}

OpenThai (formerly OpenThaiGPT) — free, open-weight Thai large language models from AIEAT and iApp Technology. With thanks to the community members who published unofficial GGUF conversions before these existed.