typhoon-ai/llama3.2-typhoon2-t1-3b-research-preview
The typhoon-ai/llama3.2-typhoon2-t1-3b-research-preview is a 3.2 billion parameter instruct decoder-only model from SCB 10X, built on the Llama architecture. As the first in the new "Typhoon T" family, it is a novel open reasoning model designed to "think longer" before providing answers, excelling across diverse domains beyond typical math and coding. This model demonstrates improved performance on challenging benchmarks like GPQA, MMLU Pro, and AI Mathematics Olympiad, and uniquely supports generating Thai reasoning traces.
Loading preview...
Typhoon T1 3B (Research Preview) Overview
Typhoon T1 3B (Research Preview) is the inaugural model in SCB 10X's new Typhoon T family of open reasoning models. Built upon the Llama 3.2 architecture, this 3.2 billion parameter instruct model introduces a novel approach where it is designed to "think longer" before generating a final answer, enabling enhanced reasoning capabilities.
Key Capabilities & Differentiators
- Enhanced Reasoning Across Domains: Unlike many reasoning models limited to mathematics or coding, Typhoon T1 3B is capable of reasoning across various domains.
- Improved Benchmark Performance: It shows significant performance gains on challenging benchmarks such as GPQA, MMLU Pro, and the AI Mathematics Olympiad validation set compared to its base model, Typhoon 2 3B Instruct.
- Structured Thinking Paradigm: The model utilizes a new "structured thinking" paradigm with auxiliary tokens to guide its thought process, leading to increased performance.
- Multilingual Reasoning: The
v2025-02-01update specifically enables the generation of Thai reasoning traces, improving transparency and interpretability for Thai language tasks, alongside enhanced general Thai performance and instruction following. - Low-Compute, High Capability: It offers a fast model with low computational requirements, capable of scaling test-time compute to achieve robust performance.
Performance Highlights
Typhoon T1 3B (Research Preview) demonstrates superior results:
- GSM8K (8-shot): 62.40 (vs. 56.63 for Typhoon 2 3B Instruct)
- HumanEval+ (Pass@10): 69.87 (vs. 66)
- GPQA (0CoT): 31.7 (vs. 27.01)
- MMLU Pro Average (5-shot): 30.65 (vs. 26.7)
This model is ideal for applications requiring robust reasoning, especially in scenarios where a smaller, efficient model needs to tackle complex problems or provide transparent, step-by-step thought processes in both English and Thai.