Microsoft ResearchLocal time Tuesday announced its Phi-2 Small Language Model(SML)This isA text-to-text artificial intelligenceprogram, which, according to a post on the X platform, it"Small enough to run on a laptop or mobile device.".
With 2.7 billion parameters, Phi-2 rivals the performance of other larger models, including Meta's Llama 2-7B(with 7 billion parameters) and Mistral-7B (another 7 billion parameter model).
Microsoft Research has also released in its Phi-2Blog PostsIt is noted that although Phi-2 has fewer parameters than Google's brand new Gemini Nano 2 model (which has more than 5 billion parameters), but it performs better and has less 'toxicity' and bias in its responses than Llama 2.
Microsoft also took the opportunity to take a little dig at Google's release of the Gemini model, which is on the scale of its upcomingmaximumCompetenciesStrongest's new AI model, Gemini Ultra, in a demo video showing its ability to solve relatively complex physics problems and even correct student errors. It turns out that while Phi-2 may be a fraction of the size of Gemini Ultra, it can also answer questions correctly and correct students using the same prompts.
However, as encouraging as these findings are, Phi-2 currently has one major limitation: according to the Microsoft Research license, theIt is for "research purposes only" and not for commercial use.The license further states that Phi-2 may only be used for "non-commercial, non-revenue producing research purposes. The license further specifies that Phi-2 can only be used for "non-commercial, non-revenue generating research purposes." Therefore, companies wishing to build products based on it are out of luck for the time being.