Microsoft makes its Phi-4 small language model open-source

Microsoft's Phi-4, a small language model with 14 billion parameters, excels in generating text and solving mathematical problems, now available for open source.

Trained on a state-of-the-art cluster of 1,920 Nvidia H100 processors, Phi-4 demonstrates significant performance improvements over larger models in key benchmarks.

By implementing a decoder-only architecture, Phi-4 effectively reduces inference costs while maintaining high output quality, thanks to advanced post-training optimization techniques.

Phi-4 showcases better performance than larger models like Llama 3.3 70B in GPQA and MATH benchmarks, illustrating its efficiency in processing language tasks.

Read at Techzine Global