Microsoft has introduced three new generative AIs at once — Phi 4 mini reasoning, Phi 4 reasoning, and Phi 4 reasoning plus. All of them expand the “Phi” lineup, which the company is creating for developers working with applications on low-power devices. The new models are focused on solving complex tasks where accuracy and solution verification are especially important — this is what sets them apart from previous versions.
Phi 4 mini reasoning, the most compact of the new releases, has about three and a half billion parameters and was trained on a million artificially created math problems from the startup DeepSeek. Microsoft positions it as the ideal tool for educational solutions — for example, for built-in tutoring on simple devices.
Phi 4 reasoning is a model with fourteen billion parameters, trained on “high-quality” web data and specially selected demonstrations, including experience from OpenAI. According to Microsoft, it is best suited for mathematics, natural sciences, and programming, demonstrating confident results in these areas.
Phi 4 reasoning plus is an improved version of the previous model, adapted for even more accurate problem-solving in specific tasks. Microsoft’s internal tests show that in terms of mathematical skills, this AI rivals even such powerful solutions as o3-mini and approaches the results of the R1 model with a much larger number of parameters.
All three models are already available on the Hugging Face platform along with detailed technical descriptions. Microsoft emphasizes that these AIs combine compactness with powerful analytical capabilities, enabling complex computations even on devices with limited resources.
