What is TinyR1-Preview
TinyR1-Preview is a 32B-parameter reasoning model jointly launched by the School of Computer Science at Peking University and 360. Using only 5% of the parameter count, the model approaches the performance of Deepseek-R1-671B. In the field of mathematics, TinyR1-Preview (78.1 points on the AIME benchmark) approaches the original R1 (79.8 points), far exceeding the 70B Deepseek-R1-Distill-Llama (70.0 points). TinyR1-Preview is based on a 'divide-and-conquer + fusion' strategy, training separate models for the three major vertical domains of mathematics, programming, and science, and using the Mergekit tool to achieve intelligent fusion, breaking through performance limits.
Main features of TinyR1-Preview
- Powerful mathematical reasoning ability: Performs excellently on complex mathematical problems (such as AIME 2024), quickly and accurately solving highly difficult math problems.
- Efficient programming assistance: Supports code generation and debugging, helping developers solve problems quickly and improving programming efficiency.
- Scientific question answering: Supports handling complex scientific problems, providing accurate answers and explanations.
- Lightweight deployment: Requires only 32B parameters. Compared with large models, it has lower inference costs and is suitable for resource-constrained scenarios.
Technical principles of TinyR1-Preview
- Divide-and-conquer strategy: Based on massive domain data generated by DeepSeek-R1, it separately trains sub-models in vertical domains such as mathematics, programming, and science, with each sub-model focusing on tasks in a specific domain.
- Intelligent fusion: Based on the Mergekit tool from the Arcee team, it intelligently fuses sub-models from different domains, breaking through the performance limits of a single model and achieving balanced optimization across multiple tasks.
- Distillation technology: Based on model distillation methods, it transfers the knowledge of large models to smaller models, achieving more than 95% of the performance of the original R1 model with only 5% of the parameter count.
- Optimized training: Based on domain data training and intelligent fusion, TinyR1-Preview significantly improves inference efficiency and performance while maintaining lightweight characteristics, making it suitable for rapid deployment and application.
Project address of TinyR1-Preview
- HuggingFace model library：https://huggingface.co/qihoo360/TinyR1-32B-Preview
Application scenarios of TinyR1-Preview
- Education: Assists mathematics learning and programming education, providing problem-solving ideas and code generation.
- Scientific research and academia: Helps researchers answer scientific questions, design experiments, and analyze data.
- Software development: Generates code, optimizes algorithms, and improves development efficiency.
- Enterprise applications: Supports data analysis and process optimization, assisting enterprise decision-making.
- Personal life: As an intelligent assistant, provides knowledge queries and learning support.