Following the recent launch of the natural language model Phi-13 with only 1 billion parameters, Microsoft announced earlier that it would increase the number of parameters to 27 billion. The performance is comparable to the 2 billion parameter version of Meta Llama 70 and the 70 billion parameter version of Mistral.Phi-2, and can be used for various device deployment needs.
Microsoft emphasizes that Phi-2 is only 2% the size of the 70 billion parameter version of Llama 38, but it can achieve almost the same performance, even better than Google's recently announcedGemini NanoIt has higher application performance and faster execution efficiency in multi-step reasoning.
Phi-2 was trained over 96 days using 100 NVIDIA A14 accelerators. No instruction fine-tuning or manual adjustments have been performed yet. It can mainly be used in smartphones or laptop-type mobile devices.
However, Phi-2 is currently only available for use in specific research needs and is not yet open to commercial use.
