Microsoft’s most capable new Phi 4 AI model rivals the performance of far larger systems

Microsoft Several new “Open Open” launched AI models The most skilled on Wednesday is competitive with Openai O3-mini on at least one benchmark.

All new pemissive licensed models – PHI 4 is a substantiation, PHI 4 Reasoning and PHI 4 Reasoning Plus – “Reasoning” models, they can hold more facts with complex problems. The company expands the Microsoft’s family, a “small model” family presented a fund for apps on the edge of AI developers a year ago.

PHI 4 mini rationale China has trained in about 1 million synthetic mathematics problem generated by the EU Beginner DeepSeek R1 justification model. The size of 3.8 billion parameter, PHI 4 mini is designed for educational applications, Microsoft says “Internal Tutoring” on light devices such as “internal tutoring”.

Settings are approximately a model of problem solving skills and models with more parameters generally perform better than less parameters.

A 14 billion-billion parameter model was developed using “high quality web data, as well as” Curnated Democrats “, as well as” Curnated Demonstrations from the above O3-Mini-Mini. According to Microsoft, it is best for math, science and coding applications.

As for the PHI 4 justification Plus, this is the previously missed PHI-4 model of Microsoft, adapted to a model of thinking to achieve better accuracy in certain positions. Microsoft claims that the R1 of PHI 4, the more parameter (671 billion) approached the level of performance of a model. The company’s internal assessment also adapts O3-mini in Omnimath, the PHI 4 substantiator, math skills testing.

PHI 4 Mini Rationale is available on PHI 4 RESENTALLY AND PHI 4 RESISTER PLUS AI Dev Platform embraces his face Accompanied by detailed technical reports.

“To use distillation, reinforcement learning and high quality data, these [new] Balance size and performance of models, “Microsoft wrote AA Blog Post. “They are small enough for low delay environments, still larger models protect their strongest thinking skills. This confusion allows you to effectively perform complex thinking tasks to resource-limited devices.”