Microsoft’s most capable new Phi 4 AI model rivals the performance of far larger systems
Microsoft has recently introduced a series of innovative “open” AI models that are making significant advances in the field of artificial intelligence. These new models, part of the Phi small model family, have been designed to handle complex reasoning tasks and are now seen as competitive with larger systems such as OpenAI’s o3-mini.
Introducing the Next Generation of Phi Models
The new lineup includes three distinct models: mini reasoning, reasoning, and reasoning plus. All three models are offered under a permissive license and are engineered to perform detailed fact-checking when solving intricate problems. This focus on reasoning enables even devices with limited computing power to execute demanding tasks efficiently.
Phi 4 Mini Reasoning
Mini reasoning was developed using a dataset of roughly one million synthetic math problems generated by an earlier reasoning model from a Chinese AI startup. With its 3.8 billion parameters, it is particularly suited for educational applications, such as embedded tutoring systems on lightweight devices. Generally, a higher parameter count translates to enhanced problem-solving capabilities.
Read also :
Pinterest launches new tools to fight AI slop
Phi 4 Reasoning
Built with 14 billion parameters, Phi 4 reasoning was trained using both high-quality web data and carefully curated demonstrations derived from OpenAI’s o3-mini. This model shows strong performance in math, science, and coding, making it an excellent option for developers seeking robust AI support for technical applications.
Phi 4 Reasoning Plus
Phi 4 reasoning plus builds on Microsoft’s earlier Phi-4 model by adapting it into a more efficient reasoning format. Microsoft claims that this model approaches the performance of the far larger 671 billion-parameter R1 model. Internal benchmarks have also indicated that Phi 4 reasoning plus performs on par with o3-mini in math-focused evaluations.
Read also :
JetBrains releases Mellum, an open AI coding model
Efficient AI for Low-Latency Environments
In an official blog post, Microsoft explained that the new models achieve a balance between size and performance through techniques such as distillation and reinforcement learning, coupled with high-quality data. These innovations allow the models to operate effectively in low-latency settings while maintaining strong reasoning skills.
Access and Further Information
Developers and researchers can access Phi 4 mini reasoning, Phi 4 reasoning, and Phi 4 reasoning plus on the Hugging Face AI development platform. Detailed technical reports accompany each model, providing further insights into their capabilities.
As Microsoft continues to push the envelope in artificial intelligence, these new Phi 4 models mark a significant step forward—bringing high-level reasoning to applications and devices that otherwise face resource constraints.
Read also :
Meta forecasted it would make $1.4T in revenue from generative AI by 2035
