Microsoft announced the latest addition to its Phi family of generated AI models.
Called Phi-4, Microsoft claims that it improves on several areas over its predecessor, especially math problem solving. This is partly a result of improved training data quality.
Phi-4 has very limited access as of Thursday night. It is only available for use on Microsoft’s recently launched Azure AI Foundry development platform and for research purposes only in accordance with the Microsoft Research License Agreement.
This is Microsoft’s latest compact language model, available in a parameter size of 14 billion, and competes with other compact models such as GPT-4o mini, Gemini 2.0 Flash, and Claude 3.5 Haiku. While these AI models are often faster and cheaper to run, the performance of small-scale language models has gradually improved over the past few years.
In this case, Microsoft believes Phi-4’s performance gains are due to the use of “high-quality synthetic datasets” along with high-quality datasets from human-generated content and some unspecified post-training improvements.
These days, many AI labs are looking more closely at how they can innovate around synthetic data and post-training. “We have reached a pre-training data wall,” Scale AI CEO Alexander Wang tweeted Thursday, confirming several reports on the topic over the past few weeks.
Notably, the Phi-4 is the first Phi series model to be released following the departure of Sébastien Bubeck. Bubeck, previously Microsoft’s VP of AI and a key figure in the development of the company’s Phi model, left Microsoft in October to join OpenAI.
