Microsoft has announced a new member of its Phi series generative AI models—Phi-4. Compared to previous versions, Phi-4 shows improvements in several areas, particularly in its ability to solve mathematical problems, thanks to the enhanced quality of the training data.
As of Thursday evening, access to Phi-4 is relatively limited; it is only available on Microsoft's newly launched Azure AI Foundry development platform and is restricted to research purposes under Microsoft's research licensing agreement.
Image Source Note: The image was generated by AI, with image licensing provided by Midjourney
Phi-4 is Microsoft's latest small language model, featuring 14 billion parameters. It is comparable to other small models like GPT-4o mini, Gemini2.0Flash, and Claude3.5Haiku, which typically offer faster runtime and lower costs, while the performance of small language models has been steadily improving over the past few years.
Microsoft attributes the performance enhancement of Phi-4 to the use of "high-quality synthetic datasets" as well as high-quality datasets derived from human-generated content, along with undisclosed training improvements.
Many AI laboratories are currently closely monitoring the potential of synthetic data and post-training in enhancing model performance. Alexandr Wang, CEO of Scale AI, mentioned in a tweet on Thursday: "We have reached a pre-training data bottleneck." This statement also confirms some recent reports on the topic.
Additionally, it is noteworthy that Phi-4 is the first Phi series model launched after the departure of Microsoft's AI Vice President, Sébastien Bubeck. Bubeck played a significant role in Microsoft's AI domain and was a key figure in the development of the Phi models. He left Microsoft in October to join OpenAI.
