Microsoft has announced a brand new AI model called Phi-4, which is a small language model (SLM) in contrast to the large language models (LLM), that chatbots like ChatGPT and Copilot use. As well as being lightweight, Phi-4 excels at complex reasoning which makes it perfect for math and language processing.

Microsoft has released a set of benchmarks showing Phi-4 outperforming even large language models like Gemini Pro 1.5 on math competition problems.

Breakthroughs in post-training

Small language models, like ChatGPT-4o mini, Gemini 2.0 Flash, and Claude 3.5 Haiku tend to be faster and cheaper to run compared to large language models,. However, their performance has increased dramatically with recent versions.

For Microsoft, these improvements were made possibly through breakthroughs in training Phi-4 on high-quality synthetic data sets and post-training innovations. Since the bottleneck for improving AI ability has always been the vast amount of processing power and data required for the training (sometimes called the ‘pre-training data wall’), AI companies have instead been looking at ways to improve the post-training development to improve performance.

Phi-4 is currently available on Azure AI Foundry, a platform for developers to build generative AI applications. So, while Phi-4 is available under a Microsoft research license agreement, you can’t simply start chatting with it, as you would with Copilot or ChatGPT. Instead, we'll have to wait and see what people produce with it in the future.

You might also like...