Microsoft has unveiled its new line of compact AI models, the Phi 4 series, as reported by TechCrunch on Thursday.
The newly released models – Phi-4-mini-reasoning, Phi-4-reasoning, and Phi-4-reasoning-plus – are all designed as reasoning models to tackle complex problem-solving tasks.
According to Microsoft, the Phi-4-mini-reasoning model has been trained on 1 million math problems generated by Chinese AI startup DeepSeek’s R1 model. With 3.8 billion parameters, it’s optimized for educational applications.
The Phi-4-reasoning model boasts 14 billion parameters and has been trained using high-quality web data and OpenAI’s o3-mini model. It excels in math, science, and coding tasks.
As for the Phi-4-reasoning-plus, it demonstrates performance comparable to the R1 model with its 671 billion parameters. In internal tests, it showed results similar to OpenAI’s o3-mini.
Microsoft emphasizes that these new models are engineered to push beyond the limitations of compact AI systems. Through the use of high-quality data and reinforcement learning, the tech giant claims they can now compete with larger models.
Developers and researchers can access these models through the AI development platform Hugging Face.