Microsoft’s new Phi 4 AI model, which is the most powerful AI model Microsoft has ever produced, rivals the performance and capabilities of much larger systems

Microsoft on Wednesday Launched several new “open AI models”of which the most capable is competitive with OpenAI’s o3 mini on at least one benchmark.

All of the new permissively-licensed models — Phi4 mini reasoning, Phi4 reasoning, and Phi4 reasoning plus — are reasoning models. This means they can spend more quality time fact-checking complex problems. Microsoft’s Phi family of “small models” was launched by the company a year earlier to provide a foundation for AI app developers developing apps at the edge.

The R1 reasoning model of Chinese AI startup DeepSeek was used to train the Phi 4 mini reasoning. Phi 4 mini reasoning, which has 3.8 billion parameters, is designed to be used in educational applications. Microsoft describes it as “embedded tutorials” on lightweight devices.

The number of parameters roughly corresponds to the model’s ability to solve problems. Models with more parameters perform better than models with fewer parameters.

The Phi 4 reasoning model, which has 14 billion parameters, was trained with “high-quality web data” and “curated demonstrations” taken from OpenAI’s mentioned o3 mini. Microsoft says it’s best suited for math, science, and coding applications.

Phi 4 reasoning is Microsoft’s Phi 4 model, adapted to a reasoning model for better accuracy in certain tasks. Microsoft claims that Phi 4 reasoning + is close to DeepSeek R1, a model with significantly more parameters (671 trillion). In the company’s internal benchmarking, Phi 4 reasoning and plus are both able to match o3-mini in OmniMath a math skills assessment.

The detailed technical reports for Phi mini reasoning, the Phi reasoning, and Phi reasoning plus tests, as well as their detailed technical reports are available at the AI dev platform Hugging Face.

These [new] [models]balance size and performance using distillation, reinforcement-learning, and high quality data,” wrote Microsoft. blog post. “They are small for low-latency environment yet maintain strong reasoning abilities that rival models much larger.” This blend allows even devices with limited resources to perform complex reasoning tasks effectively.

View Bio