Microsoft is expanding its Phi family with a new model, Phi-4. This small language model is said to be suitable for mathematical problems and has 14 billion parameters.
The newest language model in the Phi family is known: Phi-4. This small language model excels at complex reasoning tasks such as mathematical problems, according to a Microsoft blog post. The model is trained on “high-quality synthetic datasets” and is said to outperform even models including Llama, GPT-4o or Gemini Pro 1.5. Phi-4 is available on Azure AI Foundry.
Mathematical reasoning
Microsoft’s new Phi-4 language model is the successor to Phi-3-mini, which had only 3.8 billion parameters. For Phi-4, Microsoft scaled up with as many as 14 billion parameters. The model is said to excel primarily in solving mathematical problems and score better than alternative language models, including Llama, GPT-4o and even Gemini Pro 1.5 or Claude 3.5 Sonnet.
“Phi-4 outperforms comparable and larger models in mathematical reasoning due to advances in processes, including the use of high-quality synthetic datasets, curation of high-quality organic data, and post-training innovations. Phi-4 continues to push the limit of size versus quality,” Microsoft said. The company does not reveal more details than that it is trained on “high-quality synthetic datasets. The model is available on Azure AI Foundry.
read also