Microsoft Research has released a small language AI model called Phi-2. This model boasts 2.7 billion parameters and performs comparably to larger models like Llama 2-7B and Mistral-7B. Compared to Google's Gemini Nano 2 model, Phi-2 offers better performance with less toxicity and bias. Microsoft also took a jab at Google's Gemini model, demonstrating Phi-2's ability to correctly answer questions and correct student mistakes. However, Phi-2 is currently limited to research purposes and cannot be used for commercial applications.