Microsoft Unveils Phi 4 AI Models for Edge Computing

Microsoft has released three new open AI models, Phi 4 mini reasoning, Phi 4 reasoning, and Phi 4 reasoning plus, which are competitive with OpenAI's o3-mini on at least one benchmark. Phi 4 reasoning plus approaches the performance levels of R1, a model with significantly more parameters, and matches o3-mini on OmniMath, a math skills test. The models were trained using synthetic math problems, high-quality web data, and curated demonstrations. They are available on the AI dev platform Hugging Face.

Ask Aime: "Are Microsoft's new AI models Phi 4 mini, Phi 4 reasoning, and Phi 4 reasoning plus outperforming OpenAI's o3-mini?"

Microsoft has recently introduced three new open AI models, Phi-4 mini reasoning, Phi-4 reasoning, and Phi-4 reasoning plus, which are competitive with OpenAI's o3-mini on at least one benchmark. These models were designed to excel in mathematical reasoning and are available on the AI development platform Hugging Face.

The models were trained using synthetic math problems, high-quality web data, and curated demonstrations. Phi-4 reasoning plus approaches the performance levels of R1, a model with significantly more parameters, and matches o3-mini on OmniMath, a math skills test. This indicates that smaller models can achieve high levels of performance when trained effectively.

The Phi-4 mini reasoning model is particularly notable for its efficiency, with a size of just 3.8 billion parameters. It supports a context length of 128K tokens, making it suitable for environments with constrained computing or latency. This model is ideal for educational applications, embedded tutoring, and lightweight deployment on edge or mobile systems.

The Phi-4 reasoning model is a solid all-rounder, capable of handling a wide range of tasks, including math, coding, and planning. It uses custom think tags to structure logical blocks and has a context window of 32K tokens, allowing it to process longer documents.

The Phi-4 reasoning plus model builds upon the capabilities of the base model, using reinforcement learning to deliver higher accuracy. It is particularly strong in mathematical reasoning and can handle longer reasoning traces, although this comes at the cost of increased inference time.

These models are part of Microsoft's ongoing effort to advance and democratize artificial intelligence through open source and open science. They are designed to be small and efficient, making them accessible to a wide range of users and devices.

References:

[1] https://medium.com/data-science-in-your-pocket/phi-4-reasoning-microsofts-new-llms-are-smarter-faster-free-er-a477e832aae8
[2] https://azure.microsoft.com/en-us/blog/one-year-of-phi-small-language-models-making-big-leaps-in-ai/

View Source