On Friday, Microsoft launched its Phi-4 artificial intelligence (AI) model, expanding its lineup of small language models (SLMs) within the Phi open-source family. This release comes eight months after the launch of Phi-3 and four months following the introduction of Phi-3.5. The company asserts that the new model enhances capabilities in solving intricate reasoning tasks, particularly in mathematics, while also demonstrating improvements in traditional language processing tasks.
Microsoft’s Phi-4 AI Model to Be Available via Hugging Face
Traditionally, each release in the Phi series has included a mini variant, but Phi-4 does not have a mini equivalent at this time. According to a blog post by Microsoft, the Phi-4 model is currently accessible on Azure AI Foundry under the Microsoft Research Licence Agreement (MSRLA), with plans to become available on Hugging Face in the coming week.
Microsoft also disclosed benchmark results from internal evaluations, indicating that Phi-4 represents a substantial advancement over its predecessors. The company claims that the new model surpasses Gemini Pro 1.5, a larger model, on benchmark tests involving mathematical competition problems. Detailed benchmark performance statistics have been made available in a technical paper published on the online platform arXiv.
Regarding safety features, Microsoft noted that the Azure AI Foundry includes tools designed to assist organizations in assessing, mitigating, and managing AI-related risks throughout the development process for both traditional machine learning and generative AI applications. Enterprise users can leverage Azure AI Content Safety features, including prompt shields and groundedness detection, to filter content effectively.
Developers have the option to integrate these safety features into their applications via a streamlined application programming interface (API). The platform is capable of monitoring for quality and safety, addressing adversarial prompt attacks, and ensuring data integrity, offering real-time alerts to developers. These functionalities will be accessible to users working with Phi via Azure.
It’s worth noting that smaller language models are frequently enhanced post-deployment through training on synthetic data, which allows for rapid knowledge acquisition and greater efficiency. Nevertheless, the performance of these models after training may not always align with real-world application outcomes.
