Microsoft presents Phi-4, a new AI that is more advanced in mathematics and language processing.
No, this does not replace Copilot.
Microsoft has introduced its new artificial intelligence model, Phi-4, which is categorized as a small language model (SLM), setting it apart from the large language models (LLMs) used by chatbots like ChatGPT and Copilot. Despite its smaller size, Phi-4 excels in complex reasoning tasks, making it an ideal choice for solving mathematical problems and language processing.
The company has shared a series of metrics demonstrating that Phi-4 outperforms even larger models like Gemini Pro 1.5 in competitive mathematics problems. These smaller models, such as ChatGPT-4o mini, Gemini 2.0 Flash, and Claude 3.5 Haiku, are generally faster and more cost-effective compared to their larger counterparts, and their performance has significantly improved with recent versions.
For Microsoft, these advancements were achieved through innovations in training Phi-4 using high-quality synthetic datasets, as well as post-training developments. Traditionally, the main hurdle to enhancing AI capability has been the massive amount of processing power and data needed for training, referred to as the "pre-training data wall." In this regard, several AI companies are exploring ways to improve post-training development to optimize performance.
Phi-4 is already available on Azure AI Foundry, a platform designed for developers to create generative AI applications. However, while Phi-4 can be accessed under a Microsoft research license agreement, it cannot be interacted with in the same way as Copilot or ChatGPT, leaving anticipation about what developers will be able to create with this new model in the future.
