The Phi-3 Mini, the first of three lightweight models from Microsoft, is the company’s smallest AI model to date.
Microsoft is exploring models that are trained on smaller-than-usual datasets as an increasing number of AI models enter the market. According to The Verge, Phi-3 Mini is now available on Hugging Face, Ollama, and Azure. It has 3.8 billion parameters, or the number of complex instructions a model can understand. Two more models are planned for release. Phi-3 Medium and Phi-3 Small measure 14 billion parameters and seven bullion parameters, respectively. It is estimated that ChatGPT 4 contains more than a trillion parameters, to put things into perspective.
Released in December 2023, Microsoft’s Phi-2 model has 2.7 billion parameters and can achieve performance levels comparable to some larger models. According to the company, Phi-3 can now perform better than its predecessor, providing responses that are comparable to those that are ten times larger.
Benefits of the Phi-3 Mini
Generally speaking, smaller AI models are less expensive to develop and operate. Because of their compact design, they work well on personal computers and phones, which facilitates their adaptation and mass market introduction.
Microsoft has a group devoted to creating more manageable AI models, each with a specific focus. For instance, as its name would imply, Orca-Math is primarily concerned with solving math problems. T.
There are other companies that are focusing on this field as well. For example, Google has Gemma 2B and 7B that are focused on language and chatbots, Anthropic has Claude 3 Haiku that is meant to read and summarize long research papers (just like Microsoft’s CoPilot), and Meta has Llama 3 8B that is prepared to help with coding.
Although smaller AI models are more suitable for personal use, businesses may also find use for them. These AI models are ideal for internal use since internal datasets from businesses are typically smaller, they can be installed more quickly, are less expensive, and easier to use.