openthaigpt/openthaigpt-r1-32b-instruct
OpenThaiGPT R1 32b is a 32-billion-parameter Thai language reasoning model developed by OpenThaiGPT. It is specifically designed for complex reasoning tasks, including mathematics, logic, and code reasoning in Thai, outperforming larger 70-billion-parameter models like DeepSeek R1 70b and Typhoon R1 70b in these specialized areas. This model provides explicit step-by-step thought processes and demonstrates high performance on code reasoning in both Thai and English.
Loading preview...
OpenThaiGPT R1 32b: Advanced Thai Reasoning Model
OpenThaiGPT R1 32b is a 32-billion-parameter model developed by OpenThaiGPT, specifically engineered for advanced reasoning in the Thai language. It distinguishes itself by outperforming larger 70-billion-parameter models, such as DeepSeek R1 70b and Typhoon R1 70b, in complex reasoning tasks while being less than half their size.
Key Capabilities
- Superior Thai Reasoning: Excels in mathematical, logical, and code reasoning tasks in Thai.
- Explicit Step-by-Step Thinking: Capable of demonstrating detailed thought processes for problem-solving.
- Efficient Performance: Achieves high performance with a significantly smaller parameter count (32B) compared to other models in its class.
- Multilingual Code Reasoning: Performs well in code reasoning for both Thai and English.
Benchmark Highlights
OpenThaiGPT R1 32b shows strong performance across various benchmarks, particularly in Thai-centric reasoning:
- AIME24-TH: 56.67 (vs. DeepSeek R1 70b: 33.33, Typhoon R1 Distill 70b: 53.33)
- MATH500-TH: 83.8 (vs. DeepSeek R1 70b: 75.4, Typhoon R1 Distill 70b: 81)
- LiveCodeBench-TH: 62.16 (vs. DeepSeek R1 70b: 53.15, Typhoon R1 Distill 70b: 47.75)
Its average score across these benchmarks is 71.58, surpassing the 63.31 of DeepSeek R1 70b and 65.42 of Typhoon R1 Distill 70b.
When to Use This Model
This model is ideal for applications requiring robust Thai language reasoning, especially in domains like education, technical support, or any scenario demanding precise logical and mathematical problem-solving in Thai. Its smaller size also makes it more efficient for deployment on systems with limited GPU resources, requiring approximately 16 GB RAM/VRAM for 4-bit quantized versions.