Microsoft has publicly released the most recent member of its Phi family of generative artificial intelligence models, Phi-4. Microsoft argues that the model is superior to its predecessors in several areas, particularly its ability to solve mathematical problems. The improved quality of the training data is another contributing factor.
As of Thursday night, Phi-4 was only accessible through a restricted access system on Microsoft’s Azure AI Foundry development platform, which was introduced on November 19. It is only available through a research licence agreement with Microsoft.
Phi-4’s size and competitors
This is the most recent small language model that Microsoft has developed. It has a size of 14 billion parameters and will compete with other small models such as the GPT-4o micro, Gemini 2.0 Flash, and Claude 3.5 Haiku. The performance of these smaller artificial intelligence models has gradually improved throughout the past few years, and they are frequently faster and less expensive to run.
Read also: Google’s Gemini 2.0 Flash: AI tool that mimics playful voice, creates and edits images, texts
Microsoft’s Phi-4 performance enhancements
In this particular instance, Microsoft attributed the boost in performance of Phi-4 to the utilisation of “high-quality synthetic datasets” in conjunction with high-quality datasets of human-generated material, as well as some specific post-training enhancements that are not stated. These days, many artificial intelligence labs focus on the advancements they can make concerning synthetic data and post-training. Alexandr Wang, CEO of Scale AI, said in a tweet on Thursday that “we have reached a pre-training data wall.” This statement adds to the numerous reports published over the past few weeks.
Significance of Phi-4’s release
Particularly noteworthy is that the Phi-4 is the first model in the Phi series to be released after Sébastien Bubeck left Toyota. Bubeck, who previously served as one of the vice presidents of artificial intelligence at Microsoft and played a pivotal role in developing the company’s Phi model, left in October to join OpenAI.