On Wednesday, Microsoft launched a new lineup of “open” AI models, adding to its Phi model family. The standout in this release is a model that competes with OpenAI’s o3-mini, excelling in at least one benchmark. These new models ā Phi 4 mini reasoning, Phi 4 reasoning, and Phi 4 reasoning plus ā focus on āreasoning,ā meaning they spend more time fact-checking solutions to complex problems.
Introducing the Phi 4 Family
The Phi 4 models are designed to cater to AI developers working with apps that run at the edge. Microsoft first introduced the Phi model family a year ago, aiming to provide small, efficient models for various applications. These latest models are all about balancing size with reasoning capabilities, allowing them to deliver powerful performance while staying small enough for lower-latency environments.
Phi 4 Mini Reasoning: Small But Powerful
The Phi 4 mini reasoning model is a compact, 3.8-billion-parameter AI trained on about 1 million synthetic math problems created by Chinese AI startup DeepSeekās R1 model. The Windows maker has emphasised that this version is made for learning purposes. Its small size and specific training set make it particularly useful for educational applications.
Phi 4 Reasoning: A Stronger Model for STEM Applications
Phi 4 reasoning ups the ante with 14 billion parameters, drawing on high-quality web data and curated examples from OpenAIās o3-mini. Microsoft says this model is excellent in areas like math, science and coding. This would make it an essential tool for developers in such areas.
Phi 4 Reasoning Plus: Performance at Par with Larger Models
The Phi 4 reasoning plus model is an adapted version of the original Phi 4 model, re-engineered for enhanced accuracy on specific tasks. Microsoft claims this model can rival the performance of models with much more computing power, including the R1 model with 671 billion parameters. In internal benchmarks, Phi 4 reasoning plus matched OpenAIās o3-mini on a math test known as OmniMath, demonstrating impressive accuracy.
Accessible for Developers
All three models are now available for developers via the AI platform Hugging Face. This is the platform. where they come with detailed technical documentation. Microsoft also highlighted that the new models are the result of combining distillation, reinforcement learning, and high-quality data. This ensures they offer strong reasoning abilities without requiring massive computational resources. This balance makes them ideal for resource-limited devices while still being capable of handling complex tasks.
