☁️ SiamGPT-32B: A Robust Spokesperson for Linguistically Stable Thai Generation

🇹🇭 SiamGPT-32B is a fine-tuned variant of Qwen3-32B, optimized specifically for high-fidelity Thai language generation, instruction following, and multi-turn dialogue stability. Developed by SiamGPT, this model addresses the critical challenge of "multilingual interference" (code-switching) often found in open-weight models when processing Thai.

ℹ️ Model Description

• Model type: A 32B instruct decoder-only model based on Qwen3 architecture.
• Primary Language(s): Thai 🇹🇭 and English 🇬🇧
• Context Length: 8,192 tokens
• License: Apache 2.0 License

🏆 Achieves the highest overall score among open-weights models in the 30B-32B class on the SEA-HELM Thai benchmarks.

🚀 Key Features

• 🗣️ Reduced Code-Switching: Specifically trained to mitigate the injection of Chinese, Hindi, or English tokens into Thai sentences, ensuring output suitable for production user-facing applications.
• 🧠 Agentic Focus: Designed as a final response synthesizer for multi-agent systems, prioritizing strict formatting constraints and reasoning over open-ended creative writing.

🥇 SEA-HELM Leaderboard (Thai)

Full comparison against regional models (Typhoon 2.5 & OpenThaiGPT)

Metric	SiamGPT-32B (Ours)	Typhoon 2.5	OpenThaiGPT R1
Total Overall	63.59	60.44	55.28
Instruction Following	83.00	79.00	54.00
Multi-turn	75.81	76.16	59.69
NLU (Understanding)	67.95	65.56	59.89
NLG (Generation)	42.06	56.70	54.31
NLR (Reasoning)	68.59	55.54	65.38
Safety	44.19	29.68	41.42

📈 Stability & Control vs. Base Model

Full ablation analysis against Qwen3-32B baseline

Metric	Qwen3-32B	SiamGPT-32B	Improvement
Stability (Code Switch)	87.70	90.40	+2.70
Instruction Following (IF-Eval)	75.47	83.00	+7.53
Multi-Turn Dialogue (MT-Bench)	57.94	75.81	+17.87
Thai Exam (Knowledge)	61.40	63.00	+1.60
NLU (Natural Language Understanding)	59.80	67.95	+8.15

💻 Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "siamaids/SiamGPT-32B"

tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    device_map="auto",
    torch_dtype="auto"
)

messages = [
    {"role": "system", "content": "You are SiamGPT, a helpful Thai AI assistant."},
    {"role": "user", "content": "ช่วยแนะนำสถานที่เที่ยวในเชียงใหม่ให้หน่อย"}
]

text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)

generated_ids = model.generate(
    **model_inputs,
    max_new_tokens=512
)
generated_ids = [
    output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
]

response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
print(response)

⚡️ Deploy via vLLM

• vllm >= 0.8.5

pip install vllm
vllm serve siamaids/SiamGPT-32B --max-model-len 8192 --reasoning-parser qwen3 --gpu-memory-utilization 0.95

💥 Deploy via SGLang

• sglang >= 0.4.6post1

pip install sglang
python -m sglang.launch_server --model-path siamaids/SiamGPT-32B --reasoning-parser qwen3

⚠️ Limitations & Risks

• "Translationese" Artifacts: Due to the heavy reliance on translated instruction data, the model may occasionally exhibit phrasing that, while grammatically correct, lacks the stylistic naturalness of native Thai speakers.
• Looping in Creative Writing: The model is optimized for agentic, instruction-following tasks.When used for open-ended creative writing without strong context anchoring, it may exhibit repetition or looping.
• Factuality: Like all LLMs, SiamGPT can hallucinate.It is recommended for use as a response synthesizer in RAG (Retrieval-Augmented Generation) pipelines where ground truth is provided in the context.

📚 Citation

If you find SiamGPT useful for your work, please cite it using:

@misc{pairatsuppawat2025siamgptqualityfirstfinetuningstable,
      title={SiamGPT: Quality-First Fine-Tuning for Stable Thai Text Generation}, 
      author={Thittipat Pairatsuppawat and Abhibhu Tachaapornchai and Paweekorn Kusolsomboon and Chutikan Chaiwong and Thodsaporn Chay-intr and Kobkrit Viriyayudhakorn and Nongnuch Ketui and Aslan B. Wong},
      year={2025},
      eprint={2512.19455},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={[https://arxiv.org/abs/2512.19455](https://arxiv.org/abs/2512.19455)}, 
}

🤝 Connect With Us

• 🌐 SiamGPT Official: siamgpt.com

🙏 Acknowledgements

We extend our sincere gratitude to everyone using SiamGPT, providing feedback, and building with our models.

SiamGPT is an AI research and product company dedicated to advancing frontier language models for Thai and regional languages. Our training, fine-tuning, and high-throughput inference are powered by the high-performance GPU-as-a-Service infrastructure from SIAM.AI CLOUD.

Every interaction helps us refine our models and push Thailand's AI capabilities forward. Thank you for being part of this journey! 🇹🇭 ✨

☁️🇹🇭 SiamGPT Powered by SIAM.AI CLOUD