Model Overview

amd/ReasonLite-0.6B-Turbo is an ultra-lightweight 0.6 billion parameter model specifically designed for math reasoning. Developed by AMD, this model is a result of a two-stage progressive distillation process, initially from Qwen3-0.6B, using 6.1 million high-quality samples. It is notable for extending the scaling law of small models, achieving performance levels typically seen in models over 10 times its size, such as Qwen3-8B.

Key Capabilities

Exceptional Math Reasoning: Achieves 57.1 on AIME24, significantly improving over its base model (Qwen3-0.6B at 11.0).
Ultra-Lightweight: With only 0.6B parameters, it offers high efficiency for deployment in resource-constrained environments.
Distilled Performance: Utilizes a two-stage distillation process with short-CoT data to balance performance and efficiency.
Fully Open-Source: Provides access to weights, scripts, datasets, and the synthesis pipeline.

Evaluation Highlights

ReasonLite-0.6B-Turbo demonstrates strong performance across various math reasoning benchmarks. For instance, it scores 81.6 on AMC23 avg@16 and 42.7 on AIME25 avg@16, showcasing its capability in complex mathematical problem-solving despite its small size.

Good For

Applications requiring efficient and accurate mathematical reasoning.
Edge devices or environments with limited computational resources.
Developers interested in exploring high-performance, small-scale language models for specialized tasks.