Connect with us

News

Microsoft has released a Phi-2 model that “can run on a mobile device”

However, despite these encouraging results, the Phi-2 has a serious limitation, at least for now – the model is only licensed for “research purposes” and not for commercial use.

The rapid pace of releasing more and more generative AI models isn’t slowing down, even as we approach the 2023 finals and the traditional winter holiday lull.

Microsoft Research today announced Phi-2, a small language model (SML) for word processing that is “small enough to run on a laptop or mobile device,” the company said.

At the same time, Phi-2, with its 2.7 billion parameters (connections between artificial neurons), boasts performance comparable to other, much larger models, including Meta*’s Llama 2 7B with 7 billion parameters and even Mistral-7B, another model with 7 billion parameters.

Microsoft researchers also noted in their blog post on the Phi-2 release that it outperforms Google’s new Gemini Nano model despite the fact that the latter has half a billion more parameters and provides less “toxicity” and response bias than the Llama 2.

Microsoft also couldn’t resist poking fun at Google, which is now being heavily criticised for its staged Gemini demonstration video, in which the company showed how its upcoming Gemini Ultra, its biggest and most powerful new AI model, can solve quite complex physics problems and even correct students’ mistakes in them. As it turns out, despite the fact that Microsoft’s model is likely many times smaller than the Gemini Ultra in terms of size, Phi-2 was also able to answer the question correctly and correct the student using the same hints.

Microsoft выпустила модель Phi-2, которая "может работать на мобильном устройстве"

However, despite these encouraging results, the Phi-2 has a serious limitation, at least for now – the model is only licensed for “research purposes” and not for commercial use, according to the Microsoft Research Special Licence, which also states that the Phi-2 can only be used for “non-commercial, non-revenue generating research purposes”. Thus, companies wishing to build products based on it are out of luck for now.

Advertisement

Trending