Microsoft introduced Phi-4, a brand new AI that’s higher at math and language processing

Microsoft declares new AI mannequin Phi-4
It is out there to builders and researchers now
Performs nicely at math duties regardless of its small scale

Microsoft has introduced a model new AI mannequin referred to as Phi-4, which is a small language mannequin (SLM) in distinction to the massive language fashions (LLM), that chatbots like ChatGPT and Copilot use. In addition to being light-weight, Phi-4 excels at advanced reasoning which makes it excellent for math and language processing.

Microsoft has launched a set of benchmarks displaying Phi-4 outperforming even massive language fashions like Gemini Professional 1.5 on math competitors issues.

microsoft phi-4 benchmarks — (Picture credit score: Microsoft)

Breakthroughs in post-training

Small language fashions, like ChatGPT-4o mini, Gemini 2.0 Flash, and Claude 3.5 Haiku are usually quicker and cheaper to run in comparison with massive language fashions,. Nonetheless, their efficiency has elevated dramatically with current variations.

For Microsoft, these enhancements had been made presumably by breakthroughs in coaching Phi-4 on high-quality artificial information units and post-training improvements. For the reason that bottleneck for bettering AI capability has at all times been the huge quantity of processing energy and information required for the coaching (typically referred to as the ‘pre-training information wall’), AI firms have as an alternative been taking a look at methods to enhance the post-training improvement to enhance efficiency.

Phi-4 is at present out there on Azure AI Foundry, a platform for builders to construct generative AI functions. So, whereas Phi-4 is offered below a Microsoft analysis license settlement, you’ll be able to’t merely begin chatting with it, as you’ll with Copilot or ChatGPT. As a substitute, we’ll have to attend and see what folks produce with it sooner or later.