Microsoft recently updated the license of its 2.7 billion parameter small model "Phi-2," which was released in December, to a more permissive MIT open source license. This license allows for wide-ranging permissions, including commercial use, distribution, modification, and private use, with the only requirement being that users must retain copyright and license statements. However, Phi-2, as an artificial intelligence model, is provided "as is" without any form of warranty, and developers are still free to use and modify it according to their needs.

On December 12th of last year, Microsoft released more details about Phi-2, including detailed benchmark tests comparing it to the mysterious 7B model, Llama-2, and Google's Gemini Nano model. In all benchmark tests showcased by Microsoft, the performance of the 270 million parameter Phi-2 surpassed that of Google's smallest language model, Gemini Nano.

Microsoft also conducted extensive testing of Phi-2 with common prompt phrases, and the results showed, "We observed behavior that aligns with what we would expect based on benchmark results."

It is reported that Gemini Nano will be applied to devices like Pixel 8.

In an initial report dating back to November 16, 2023, Microsoft researchers introduced Phi-1, a Transformer-based language model with only 1.3 billion parameters. This model, trained solely on high-quality data, outperformed models ten times its size in benchmark tests.

Shortly after, Phi-1.5 was released, also with 1.3 billion parameters, and underwent additional data training, including various AI-generated texts. Phi-1.5 is capable of composing poetry, writing emails and stories, and summarizing textual content. One variant even possesses the ability to analyze images. In benchmark tests of common sense, language comprehension, and reasoning, this model can rival models with up to 10 billion parameters in certain areas.

Microsoft then announced Phi-2, equipped with 2.7 billion parameters, which is still considered compact compared to other language models. Compared to Phi-1.5, the company claims that Phi-2 has made significant improvements in logical reasoning and security. Microsoft asserts that with proper fine-tuning and customization, small language models can become powerful tools for cloud and edge applications.

In addition, Microsoft announced "Model as a Service." Phi-2 and Phi-1.5 have now joined the Azure AI Model Gallery, which also includes models such as Stable Diffusion, Falcon, CLIP, Whisper V3, BLIP, and SAM. Microsoft has also added Code Llama from Meta and Nemotron from Nvidia.

The launch of "Model as a Service" means that professional developers will soon be able to easily integrate the latest AI models, such as Meta's Llama 2, into their applications. They can also fine-tune these models with their own data, eliminating the complexity of configuring resources and managing hosted infrastructure without worrying about setting up and managing GPU infrastructure.