Microsoft Research has unveiled Phi-2, its latest small language model (SML), which is notable for its ability to run on consumer-grade hardware like laptops and mobile devices. This new AI model comes at a time when the generative AI field is buzzing with activity, even as 2023 winds down. Phi-2 stands out for its compactness, consisting of 2.7 billion parameters, a size that doesn’t sacrifice performance. It competes with larger models such as Meta’s Llama 2-7B and Mistral-7B, both of which have 7 billion parameters.
Phi-2 isn’t just about size efficiency; it’s also about quality. Microsoft’s researchers highlight that Phi-2 surpasses Google’s Gemini Nano 2 model in performance, despite the latter having more parameters. Moreover, Phi-2 is reported to generate responses with lower levels of toxicity and bias compared to its competitors, such as Llama 2, indicating advancements not only in technical capability but also in ethical AI development.
In an apparent jab at Google, Microsoft pointed out that their Phi-2 model could solve complex physics problems and correct mistakes in a manner similar to Google’s Gemini Ultra, despite Phi-2 likely being significantly smaller. This was a reference to Google’s controversial demo of Gemini, which was critiqued for its staged nature. Microsoft’s demonstration served to showcase Phi-2’s capabilities in practical problem-solving scenarios.
Despite these promising capabilities, Phi-2 is currently limited to non-commercial research applications. Under the custom Microsoft Research License, it cannot be used for commercial or revenue-generating projects. This restriction places limits on businesses that might wish to leverage Phi-2’s abilities for product development, at least for now. Microsoft’s licensing terms reinforce the model’s position as a tool for exploration and academic inquiry in the rapidly evolving field of AI.
Source: Microsoft releases Phi-2, a small language model AI that outperforms Llama 2, Mistral 7B
