Hugging Face Overview
Model Overview
The amd/ReasonLite-0.6B-Turbo is an ultra-lightweight 0.6 billion parameter model from AMD, specifically designed for mathematical reasoning. It stands out by achieving performance levels comparable to models over 10 times its size, such as Qwen3-8B, particularly on the AIME24 benchmark where it scores 57.1.
Key Capabilities & Features
- Exceptional Math Reasoning: Achieves strong performance on AIME24, demonstrating advanced mathematical problem-solving capabilities for its size.
- Ultra-Lightweight: With only 0.6 billion parameters, it offers significant efficiency advantages, making it suitable for resource-constrained environments.
- Two-Stage Distillation: Trained using a progressive two-stage distillation process, first with short-CoT data to achieve ReasonLite-0.6B-Turbo, then with long-CoT data for the higher-performing ReasonLite-0.6B variant.
- Open-Source: Fully open-source, including weights, training scripts, datasets, and the synthesis pipeline, promoting transparency and community contributions.
- High-Quality Data: Distilled using 6.1 million high-quality samples derived from 343K math problems from Polaris and OpenMathReasoning, with pseudo-labels generated via majority voting from GPT-OSS.
When to Use This Model
- Resource-Constrained Math Applications: Ideal for scenarios requiring strong mathematical reasoning capabilities on devices or platforms with limited computational resources.
- Efficient Reasoning Tasks: Suitable for integrating advanced math reasoning into applications where speed and efficiency are critical.
- Research and Development: Provides a fully open-source foundation for further research into efficient and high-performing small language models for specialized tasks.