The most capable new PHI 4 model of Microsoft rivals the performance of far larger systems

Microsoft Launched several new “open” models AI On Wednesday, the most competent of which is competitive with Openai’s O3-Mini of at least one indicator.

All the newly licensed licensed models-Phi 4 Mini Massing, PHI 4 Reflections and PHI 4 Reflections plus “Reflections” models, which means that they can take more time to check the facts about complex problems. They expand the family of Microsoft’s Small Model, which the company started a year ago to offer an AI Developers Foundation, which builds applications to build applications.

Phi 4 mini reasoning was trained in approximately 1 million synthetic mathematical problems generated by the R1 reasoning model of Chinese AI launch Deepseek. About 3.8 billion parameters in size

The parameters are roughly corresponding to the skills to solve model problems and models with more parameters are usually presented better than those with fewer parameters.

Phi 4 reasoning, a model of 14 billion parameters, was trained using “high quality” web data, as well as “cure demonstrations” by the above-mentioned O3-Mini of Openai. It is best for apps for mathematics, science and encoding, according to Microsoft.

As for the Phi 4 Massing Plus, this is the pre-issued Microsoft Phi-4 model adapted to a reasoning model to achieve better accuracy of specific tasks. Microsoft claims that Phi 4 Massioning Plus approaches R1 performance levels, a model with significantly more parameters (671 billion). The internal comparative comparison of the company also has a Phi 4 reflections plus a match of Omnimath O3-Mini, a mathematical skills test.

Phi 4 mini reasoning, phi 4 reasoning and phi 4 reasoning plus are available on AI DEV Platform to embrace face accompanied by detailed technical reports.

“Using distillation, reinforcement data and high quality data, these (new) models of balance size and performance,” writes Microsoft in a Blog postS “They are small enough for low latency environments, but maintain strong possibilities for reasoning that compete with much larger models. This mixture allows even devices with limited resources to perform effectively complex tasks for reasoning.”