Microsoft Unveils New Open-Source AI Reasoning Models: Phi
Microsoft Expands Its AI Model Family with New Reasoning Models
Microsoft has broadened its AI offerings with the introduction of new models under its Phi family. They recently launched three open-source reasoning models: Phi 4 Reasoning, Phi 4 Reasoning Plus, and Phi 4 Mini Reasoning. Among these, the standout model is designed to rival OpenAI’s o3-mini.
The Function of Reasoning Models
Reasoning models are specialized AI systems that focus on assessing facts and resolving intricate challenges. Microsoft’s latest models emphasize a blend of compact size and robust performance. According to Microsoft, their small size enables them to function effectively in low-latency settings without sacrificing the quality of their reasoning capabilities.
"They are small enough to operate in low-latency environments but retain powerful reasoning capabilities similar to that of significantly larger models," Microsoft states. This allows devices with limited resources to tackle complex reasoning tasks efficiently.
Overview of the New Phi Models
Here’s a closer look at each of the new Phi models and their unique features:
1. Phi 4 Reasoning
- Parameters: 14 billion
- Key Features:
- This open-weight model excels at generating detailed reasoning chains.
- It utilizes additional computations effectively during queries.
- The model was trained using data from the OpenAI o3-mini.
2. Phi 4 Reasoning Plus
- Parameters: Builds on Phi 4 Reasoning
- Key Features:
- This model enhances accuracy by using feedback and rewards during training.
- It requires 1.5 times more tokens for task completion, which improves precision.
3. Phi 4 Mini Reasoning
- Parameters: Approximately 3.8 billion
- Key Features:
- Designed for environments with limited resources, ideal for mobile applications.
- Trained on about one million synthetic math problems from the Chinese DeepSeek R1 model, making it suitable for educational purposes.
Performance Highlights
Microsoft claims that both Phi 4 Reasoning and Phi 4 Reasoning Plus outperform other models such as OpenAI’s o1-Mini and DeepSeek-R1-Distill-Llama-70B across a variety of benchmarks, which include challenging math reasoning and advanced Ph.D.-level inquiries. Moreover, they surpassed the larger 671-billion-parameter DeepSeek-R1 model on the AIME 2025 exam, a preliminary round for the prestigious US Mathematical Olympiad.
Availability of the New Models
The new models are accessible through Azure AI Foundry and Hugging Face, making them available for developers and researchers looking to leverage advanced reasoning capabilities in their applications.
A Brief History of Phi Models
Last December, Microsoft unveiled its Phi 4 model, a language model featuring 14 billion parameters. The primary focus of this model is on complex reasoning, especially related to mathematical problems.
With these developments, Microsoft reinforces its commitment to advancing AI technology and providing powerful tools for a range of applications, from academic research to mobile education platforms.
