Microsoft launched the preview of its “small” Phi-4 model a few days ago, which aims to mark a breakthrough. The model must compete with the biggest in the field, including Gemini 1.5 Pro. But where models like Google’s or OpenAI’s GPT-4o can handle hundreds of billions of parameters, Phi-4 is limited to 14 billion.
Microsoft widely highlights the great effectiveness of its model, with superior results in terms of mathematical “reasoning”. Performance that the publisher attributes to the use “ high quality synthetic data “. As always with model announcements, not much more is known about this data. Microsoft also mentions improvements in post-processing, but again without further details.
Phi-4 proves, according to Microsoft, to be particularly good at solving mathematical problems posed in competitions of the Mathematical Association of America. In this area, the new model would do better than the larger ones, or even their reduced versions which can be further optimized for specific needs.
The new model is currently only available in preview, with access limited to Azure AI Foundry and for research purposes only, under a license agreement. A wider launch will take place later on Hugging Face.
Phi-4 could open a new front in a battle that has until now largely focused on “bigger is better.” The power requirement and energy bill could become limiting factors, paving the way for models that aim to do more (or at least as much) for less.
