Introducing Phi-4: Microsoft’s Newest Small Language Model Specializing in Complex Reasoning

What we are going to introduce today is Φ4our state-of-the-art 14B parameter small language model (SLM), in addition to traditional language processing, is also good at complex reasoning in areas such as mathematics. Phi-4 is the latest member of our Phi family of small language models and demonstrates the possibilities as we continue to explore the boundaries of SLM. Phi-4 is currently available Azure Artificial Intelligence Foundry Available under the Microsoft Research License Agreement (MSRLA) at Face hugging next week.

Phi-4 benchmark

Phi-4 outperforms similar and larger models in mathematically relevant inference due to advances throughout the process, including the use of high-quality synthetic datasets, the management of high-quality organic datasets, and post-training innovation. Phi-4 continues to push the frontiers of size and quality.

Phi-4 is particularly good at math problems. For example, here is Phi-4’s benchmark on math competition problems:

Phi-4 Performance on Mathematics Competition Problems

On math competition problems, Phi-4 outperformed larger models, including Gemini Pro 1.5 (https://maa.org/student-programs/amc/)

To see more benchmarks, read the latest technical paper published arxiv.

Safely and responsibly implement artificial intelligence innovation

Responsibly building artificial intelligence solutions is at the core of Microsoft’s artificial intelligence development. We provide powerful, responsible AI capabilities to customers built using Phi models, Includes Phi-3.5-mini optimized for Windows Copilot+ PCs.

Azure Artificial Intelligence Foundry Provides users with a powerful set of capabilities to help organizations measure, mitigate, and manage AI risks throughout the AI ​​development lifecycle for traditional machine learning and generative AI applications. Azure AI Assessment in AI Foundry Enables developers to iteratively assess the quality and safety of models and applications using built-in and custom metrics to inform mitigation efforts.

Additionally, Phi users can use Azure AI content security Features such as prompt shielding, protected material detection and ground detection. These features can be used as content filters for any language model included in our content Model catalog Developers can easily integrate these capabilities into their applications through a single API. Once in production, developers can monitor the quality and security of their applications, adversarial real-time attacks and data integrity, and intervene promptly with real-time alerts.

Phi-4 in action

This problem demonstrates an example of mathematical reasoning that Phi-4 is capable of.

Start exploring

Phi-4 is currently available Azure Artificial Intelligence Foundrycome take a look today.