Microsoft has unveiled three new AI reasoning models—Phi-4-reasoning, Phi-4-reasoning-plus, and Phi-4-mini-reasoning. These compact language models are built for edge devices such as Windows PCs and smartphones. Among them, Phi-4-reasoning is trained on 14 billion parameters and is capable of handling advanced reasoning tasks.
Phi-4-reasoning-plus builds on the same base model as Phi-4-reasoning but leverages more inference-time compute and processes around 1.5 times more tokens to achieve greater accuracy. Despite their compact size, the Phi-4-reasoning models are competitive with significantly larger models like DeepSeek R1 671B and o3-mini.
On the GPQA benchmark, the Phi-4-reasoning-plus-14B model scores 69.3%, compared to o3-mini’s 77.7%. In the AIME 2025 test, it achieves 78%, while o3-mini reaches 82.5%. These results highlight how Microsoft’s compact model closely rivals top-tier reasoning models, despite being significantly smaller in scale.
Microsoft states that the Phi-4 reasoning models were trained using supervised fine-tuning on “carefully curated reasoning demonstrations from OpenAI’s o3-mini.” The company adds, “This model shows that with meticulous data curation and high-quality synthetic datasets, smaller models can effectively compete with much larger ones.”
In addition, the compact Phi-4-mini-reasoning model, with just 3.8 billion parameters, surpasses several 7B and 8B models. It performs competitively in benchmarks such as AIME 24, MATH 500, and GPQA Diamond, coming close to the scores of o1-mini. Microsoft notes that the Phi-4-mini was “fine-tuned using synthetic data generated by the DeepSeek-R1 model.”
Microsoft’s Phi models are already running locally on Windows Copilot+ PCs, utilizing the built-in NPU for processing. It’ll be interesting to see how the Phi-4 reasoning models enhance on-device AI performance.
