Phi-2 has 2.7 billion parameters (connections between artificial neurons) which has performance that is comparable to other, much larger models, including Meta's Llama 2-7B with its 7 billion parameters and even Mistral-7B, another 7 billion parameter model.

Writing in their bog, Microsoft boffins noted that it outperforms Google's brand new Gemini Nano 2 model despite it having half a billion more parameters and delivers less "toxicity" and bias in its responses than Llama 2.

Microsoft also couldn't resist taking a little dig at Google's now much-criticised, staged demo video for Gemini in which it showed off how its forthcoming largest and most powerful new AI model, Gemini Ultra, was able to solve pretty complex physics problems and even correct students' mistakes on them. As it turned out, even though it is likely a fraction of the size of Gemini Ultra, Phi-2 was also able to correctly answer the question and correct the student using the same prompts.

Phi-2 is limited in that it is licensed only for "research purposes only," not commercial usage, under a custom Microsoft Research License, which further states Phi-2 may only be used for "non-commercial, non-revenue generating, research purposes."  This means that businesses cannot build products on top of it.