Microsoft launches its smallest AI model yet: Phi-3 Mini

Microsoft has launched its smallest AI model yet, Phi-3 Mini, the first of a trio of lightweight models.

With more and more AI models coming onto the market, Microsoft is journeying into models that are trained on smaller-than-usual datasets. The Verge reports Phi-3 Mini measures 3.8 billion parameters (how many complex instructions a model can understand) and is now available on Azure, Hugging Face, and Ollama, with plans to release two more models. Phi-3 Small measures seven bullion parameters and Phi-3 Medium 14 billion parameters. To give a sense of scale, ChatGPT 4 is thought to have more than a trillion parameters.

Microsoft’s Phi-2 was released in December 2023, with 2.7 billion parameters but with the ability to perform at similar levels to some larger models. The company now asserts that Phi-3 can outperform its predecessor, offering responses close to those 10 times its size.

Smaller AI models are typically cheaper to both create and run. Their smaller size means they perform well on personal devices like phones and laptops, making them easier to adapt and launch into the mass market.

Microsoft has a team dedicated to developing smaller AI models, each one designed to target specific areas. Orca-Math, for example, focuses on solving math problems, as its name would suggest. T

It’s not the only company to be targeting this area, with Google’s Gemma 2B and 7B centered around language and chatbots, Anthropic’s Claude 3 Haiku designed to read and summarize lengthy research papers (similar to Microsoft’s CoPilot), and Meta’s Llama 3 8B ready to assist with coding.

While smaller AI models can work best for personal devices, there’s also potential for corporate use too. As companies’ internal datasets tend to be smaller, these AI models are well-suited for internal use, being easier and faster to install and start working with, as well as being more affordable.

Featured image: Unsplash

The post Microsoft launches its smallest AI model yet: Phi-3 Mini appeared first on ReadWrite.