The new most competent AI Phi 4 model in Microsoft competes in the performance of much larger systems

Microsoft launched several new “open” AI models Wednesday, the most capable is competitive with O3-Mini of Openai on at least one reference.

All the new models under Lomedively license – the reasoning Phi 4 Mini, the reasoning Phi 4 and the reasoning Phi 4 Plus – are models of “reasoning”, which means that they are able to pass more time warming solutions to complex problems. They extend Microsoft’s “Small Model” family, which the company launched a year ago to offer a base for AI developers to create applications to The Edge.

Phi 4 mini reasoning was formed on approximately 1 million problems of synthetic mathematics generated by the R1 model of Startup of Chinese AI Deepseek. About 3.8 billion size parameters, Phi 4 Mini reasoning is designed for educational applications, known as Microsoft, such as “integrated tutorial” on light devices.

The parameters roughly correspond to the problem solving skills of a model, and the models with more parameters generally work better than those with fewer parameters.

The Phi 4 reasoning, a model of 14 billion parameters, has been formed using “high quality” web data as well as “organized demonstrations” from O3-minini aforementioned from Openai. It is preferable for mathematics, science and coding applications, according to Microsoft.

Regarding Phi 4 reasoning, this is the previously published Phi-4 model of Microsoft adapted in a reasoning model to obtain better precision on particular tasks. Microsoft claims that the PHI 4 Plus reasoning discusses R1 performance levels, a model with many more parameters (671 billion). The company’s internal comparative analysis also has a PHI 4 reasoning more corresponding to O3-Mini on Omnimath, a skills test in mathematics.

The reasoning Phi 4 Mini, the reasoning Phi 4 and the reasoning Phi 4 Plus are available on the Ai Dev Tass Face platform accompanied by detailed technical reports.

“Using distillation, strengthening learning and high quality data, these (new) models balance size and performance,” wrote Microsoft in a blog. “They are small enough for low latency environments while keeping solid reasoning capacities that compete with much more important models. This mixture allows limited resource resources to effectively perform complex reasoning tasks. ”