Microsoft ships small but powerful new model

 

The Overview: Microsoft Research just released Phi-2, a surprisingly powerful 2.7B parameter ‘small language model’ capable of running on a laptop or phone — matching or exceeding the capabilities of larger models.

The details:

Phi-2 achieves impressive results among models under 13B parameters in reasoning, language, math, and coding benchmarks. It outdoes the 7B parameter Mistral and Llama 2 models, and even surpasses Google's 3B parameter Gemini Nano 2 on some tests. Microsoft credits strategic data selection emphasizing textbook-quality content, plus innovative knowledge transfer techniques in efficiently scaling up Phi-2. But businesses can't tap the model yet, as Microsoft currently only licenses it for non-commercial research purposes.

Why it matters: Achieving these results in such a small package is quite the feat — and the ability to run on devices like a laptop or phone opens the door for more capable local, offline usage. With the release just a week after Google unveiled Gemini Nano, Microsoft is undoubtedly making a statement!