At the Build 2024 event last night, Microsoft unveiled the compact Phi 3 complete set model after launching the Phi-3-mini in April.
The Phi 3 set includes a total of 4 models, consisting of 3 small language models (SML) and the new multimodal model. The first of Microsoft’s open-source models is the Phi-3-vision.
Phi-3-vision is a model that supports both images and text, with 4.2B parameters customized for excellent chart and map reading, deep question answering capabilities. Microsoft’s own benchmarking found that it outperformed larger models like Claude-3 Haiku and Gemini 1.0 Pro V in multiple test sets.
Individuals interested in trying out Phi-3-vision can test it through the Azure AI Studio web interface. An example of Phi-3-vision reading a chart and providing an explanation.
The language model Phi 3 also emphasizes being a compact model, using low resources to run but with high efficiency. The models introduced come in 3 sizes:
– Phi-3-mini with 3.8B parameters
– Phi-3-small with 7B parameters
– Phi-3-medium with 14B parameters
Microsoft’s demonstration showed that the mid-size model Phi-3-small with 7B parameters could outperform the much larger GPT-3.5T. Meanwhile, the top-tier Phi-3-medium model with 14B parameters surpassed Gemini 1.0 Pro.
Another intriguing aspect is that Microsoft states the Phi series models are customized to run on a variety of hardware, not just NVIDIA, but also in partnership with Intel to be compatible with Intel hardware (Xeon, Gaudi, Arc, Core Ultra) along with AMD. This broadens compatibility with popular frameworks like ONNX Runtime and DirectML for diverse applications across portable devices and web-based operations.
The Phi-3 model set is now available for use through Azure AI and Hugging Face platforms.
TLDR: Microsoft’s Phi 3 model series offers compact, high-performing language and multimodal models, customized for diverse hardware compatibility and available through Azure AI and Hugging Face.
Leave a Comment