Microsoft Unveils Phi 4 AI Models: Smaller Size, Competitive Performance with Larger Systems

On Wednesday, Microsoft unveiled several new "open" AI models, the most powerful of which can match the performance of OpenAI’s o3-mini on at least one benchmark. These new additions to Microsoft’s Phi "small model" family, which was launched a year ago, are designed to provide robust reasoning abilities, enabling them to fact-check solutions to complex problems more effectively. The models, named Phi 4 mini reasoning, Phi 4 reasoning, and Phi 4 reasoning plus, are now available on the Hugging Face platform, alongside detailed technical reports. The Phi 4 mini reasoning model was trained on approximately 1 million synthetic math problems generated by DeepSeek’s R1 reasoning model. With about 3.8 billion parameters, it is particularly suited for educational applications, such as embedded tutoring on lightweight devices. Parameters in AI models generally correlate with their problem-solving capabilities, and more parameters often lead to better performance. However, despite its relatively small size, Phi 4 mini reasoning demonstrates impressive reasoning skills, making it a valuable tool for educational purposes. The Phi 4 reasoning model, larger at 14 billion parameters, was trained using high-quality web data and curated demonstrations from OpenAI’s o3-mini. Microsoft recommends this model for applications involving math, science, and coding due to its enhanced capabilities in these areas. Phi 4 reasoning plus is an adaptation of Microsoft’s earlier Phi-4 model, modified to improve its accuracy on specific tasks. According to Microsoft, Phi 4 reasoning plus can achieve performance levels comparable to DeepSeek’s R1 model, which has a staggering 671 billion parameters. Internal benchmarks show that Phi 4 reasoning plus matches o3-mini on the OmniMath math skills test, further underscoring its efficiency and effectiveness. In a blog post, Microsoft highlighted the unique features of these models: “Using distillation, reinforcement learning, and high-quality data, these new models strike a balance between size and performance. They are compact enough to run in low-latency environments but still maintain strong reasoning capabilities that rival much larger models. This combination allows even devices with limited resources to handle complex reasoning tasks efficiently.” These advancements in AI model development offer developers a range of options for building applications that require sophisticated reasoning abilities, without the need for high computational power. By providing these smaller, yet powerful, models, Microsoft aims to make advanced AI capabilities more accessible, especially in edge computing scenarios where resources are constrained.