Microsoft Unveils Phi-2: A Compact Language Model AI Surpassing Llama 2 and Mistral 7B in Performance

The momentum of generative AI developments is accelerating as we approach the end of 2023, defying the typical slow-down associated with the winter holiday season.

Today, Microsoft Research, the cutting-edge division of the tech giant, unveiled its Phi-2 small language model (SML). This innovative text-to-text AI is compact enough to operate on laptops and mobile devices, as highlighted in a post on X.

Phi-2, equipped with 2.7 billion parameters (connections between artificial neurons), delivers performance on par with larger models like Meta's Llama 2-7B, which has 7 billion parameters, and Mistral-7B, another model of the same size. Notably, Phi-2 also excels over Google's new Gemini Nano 2 model, despite that model containing an additional half a billion parameters. Furthermore, Phi-2 generates responses with less toxicity and bias compared to Llama 2.

In a pointed comparison, Microsoft referenced Google's controversial demo video, in which the upcoming Gemini Ultra model purportedly solved complex physics problems and corrected students' mistakes. Interestingly, Phi-2 also accurately addressed these physics prompts and provided the correct solutions, despite being significantly smaller than Gemini Ultra.

However, there is a significant limitation: Phi-2 is currently licensed solely for research purposes under a custom Microsoft Research License, explicitly prohibiting commercial use. Consequently, businesses wishing to develop products based on this model will not be able to do so at this time.