Microsoft has made its newest AI models available on the Hugging Face open-source platform. These new models are Phi 4 mini reasoning, Phi 4 reasoning, and Phi 4 reasoning plus. These models are purposely built to improve reasoning capabilities, the ability to take time and fact-check answers to difficult questions. The release features in-depth technical reports describing their specifications and training process.
The new Phi 4 reasoning model has 14 billion parameters. It was further trained on high-quality web data and OpenAI-built curation demonstrations from OpenAI’s o3-mini. This model needs to provide high accuracy for specific downstream tasks, which is increasingly important in a world that depends on reliable performance in reasoning tasks. The Phi 4 mini reasoning model has approximately 3.8 billion parameters. It was specifically designed to maximize educational applications. The training consisted of about 1 million synthetic algebra problems. These issues were produced by DeepSeek’s R1 reasoning model, a Shanghai-based AI startup.
Microsoft emphasizes that these new models are small enough for low-latency environments while retaining strong reasoning capabilities that rival much larger systems. Our paradigm of design supports this kind of complex reasoning to be carried out even on devices with minimal resources.
“Using distillation, reinforcement learning, and high-quality data, these [new] models balance size and performance.” – Microsoft
The Phi 4 mini reasoning model provides outstanding “embedded tutoring” capability across compact devices. This combination of factors makes it the perfect solution when it comes to education. The flexibility of the Phi 4 reasoning plus model serves to underscore Microsoft’s goals to enhance overall AI performance. It competes with the o3-mini model on the extensive OmniMath math skills test, highlighting its ability to tackle advanced mathematical problems.
Microsoft has been intricately focused on creating smaller, but still mighty models. This coincides nicely with the industry’s ongoing push to increase efficiency and technology accessibility. The detailed reports accompanying the release provide developers with essential insights into each model’s architecture and training processes, facilitating informed integration into various applications.