On Wednesday, Microsoft unveiled a series of new open AI models, marking a significant step in the company’s AI development efforts. These new models, part of the Phi family, are designed to excel at reasoning tasks, such as fact-checking and solving complex problems. Among these, the most capable model is competitive with OpenAI’s o3-mini on at least one benchmark, pushing the boundaries of AI performance.
The newly launched models include Phi 4 mini reasoning, Phi 4 reasoning, and Phi 4 reasoning plus. Each of these models focuses on improving reasoning capabilities, a crucial aspect of AI development, especially for tasks that require in-depth understanding and validation of solutions. These models expand Microsoft’s Phi small model family, which was introduced a year ago to provide a foundational framework for AI developers building applications on edge devices.
According to Microsoft, these new models are particularly suitable for educational and resource-constrained applications. The goal is to provide powerful reasoning capabilities in small, efficient packages that can be deployed on a variety of devices, even those with limited resources.
Phi 4 Mini Reasoning: Optimized for Education
The smallest of the three new models, Phi 4 mini reasoning, is trained on about 1 million synthetic math problems generated by the DeepSeek R1 reasoning model, a product of the Chinese AI startup DeepSeek. With 3.8 billion parameters, this model is specifically designed for educational applications, such as embedded tutoring on lightweight devices.
The number of parameters in an AI model often correlates with its problem-solving ability. A higher number of parameters typically indicates that the model is capable of tackling more complex tasks. Despite being smaller than other models in the series, Phi 4 mini reasoning is designed to perform well in low-latency environments while maintaining strong reasoning capabilities. This makes it a promising solution for users who require efficient, accessible AI for tasks like math tutoring on mobile devices or other edge computing applications.
Phi 4 Reasoning: Powerful for Math, Science, and Coding
Phi 4 reasoning, the next step up, comes with an impressive 14 billion parameters. It was trained on high-quality web data and curated demonstrations, including those from OpenAI’s o3-mini, making it well-suited for math, science, and coding applications. This model is more robust than Phi 4 mini reasoning, offering greater problem-solving capabilities and versatility across a range of tasks that require advanced logic and reasoning.
This model’s large parameter size allows it to process complex information and provide solutions with a high degree of accuracy, making it a useful tool for developers who need powerful AI systems to handle technical problems. Phi 4 reasoning positions itself as an ideal choice for educational software and apps that focus on STEM subjects, where reasoning skills are key.
Phi 4 Reasoning Plus: Closely Rivals DeepSeek R1 and o3-mini
The most advanced model in the new release, builds on the Phi 4 model by optimizing it for even higher accuracy in specialized tasks. Microsoft claims that Phi 4 reasoning plus is approaching the performance levels of DeepSeek R1, which boasts an enormous 671 billion parameters. In internal benchmarks, Microsoft also reports that Phi 4 reasoning plus matches o3-mini on OmniMath, a popular math skills test.
The Phi 4 reasoning plus model represents the culmination of Microsoft’s efforts to balance size and performance in AI systems. Despite its advanced reasoning capabilities, it remains small enough to be deployed on devices with limited processing power, making it a versatile option for developers looking to bring complex reasoning tasks to a wide range of applications.
Available for Developers on Hugging Face
All three new models, Phi 4 mini reasoning, Phi 4 reasoning, and Phi 4 reasoning plus, along with their detailed technical reports, are available on the AI developer platform Hugging Face. This makes it easy for developers to integrate these advanced AI models into their applications and begin utilizing them in real-world scenarios.
Microsoft’s use of distillation and reinforcement learning techniques, combined with high-quality training data, enables these models to perform complex reasoning tasks efficiently. The models are designed to be small enough for low-latency environments, yet powerful enough to rival much larger models in terms of their problem-solving capabilities.
With the release of these new models, Microsoft is positioning itself as a strong player in the AI reasoning space. The Phi 4 reasoning models offer a unique blend of small size and high performance, making them ideal for developers working in resource-constrained environments. As AI continues to evolve, Microsoft’s focus on providing efficient, scalable solutions for edge applications could play a key role in shaping the future of AI development.
By offering these advanced reasoning models through platforms like Hugging Face, Microsoft ensures that developers have access to the latest tools for creating intelligent applications that can handle complex tasks with high accuracy.
