microsoft/Phi-4-mini-reasoning

Overview

microsoft/Phi-4-mini-reasoning is a 3.8 billion parameter model from the Phi-4 family, specifically designed for high-quality mathematical reasoning. It features a 128K token context length and is built upon synthetic data, with a strong focus on advanced math reasoning capabilities. The model was trained in February 2024 on 150 billion tokens, primarily using synthetic mathematical content generated by a more advanced reasoning model, Deepseek-R1, to distill knowledge. This dataset includes over one million diverse math problems, with verified correct solutions, totaling approximately 30 billion tokens of math content.

Key Capabilities

• Advanced Mathematical Reasoning: Excels in multi-step, logic-intensive mathematical problem-solving, including formal proof generation, symbolic computation, and advanced word problems.
• Efficiency: Optimized for memory/compute constrained environments and latency-bound scenarios, making it suitable for edge or mobile deployments.
• Context Handling: Maintains context across steps and applies structured logic for accurate, reliable solutions.
• Performance: Achieves strong results on reasoning benchmarks like AIME (57.5), MATH-500 (94.6), and GPQA Diamond (52.0), often outperforming larger models in its class.

Good For

• Mathematical Problem Solving: Ideal for tasks requiring deep analytical thinking and step-by-step logic.
• Educational Applications: Potentially suitable for embedded tutoring and similar learning tools.
• Resource-Constrained Environments: Designed for deployment where computing power or latency is a significant factor.

Limitations

While strong in reasoning, its compact size limits factual knowledge storage, potentially leading to factual incorrectness. It is primarily designed and tested for math reasoning and not evaluated for all downstream purposes. Performance may vary across languages, with English being the primary supported language.