Phi-3 is a household of open supply small language fashions developed and made obtainable by Microsoft.
“Small language fashions are designed to carry out nicely for easier duties, are extra accessible and simpler to make use of for organizations with restricted sources, and they are often extra simply fine-tuned to fulfill particular wants. They’re nicely suited to functions that have to run regionally on a tool, the place a job doesn’t require in depth reasoning and a fast response is required,” Misha Bilenko, company vp for Microsoft GenAI, wrote in a weblog publish.
The concept behind growing a mannequin so small was impressed by Microsoft researcher Ronan Elden studying a bedtime story to his daughter, which led him to suppose “how did she be taught this phrase? How does she know find out how to join these phrases?”
Making use of this to AI, Elden puzzled what would occur if an AI mannequin was skilled simply on phrases that may be understood by a 4-year-old.
Phi-3 is available in quite a lot of choices:
- Phi-3-vision is a 4.2B parameter mannequin that able to understanding each textual content and imaginative and prescient
- Phi-3-mini is a 3.8B parameter mannequin, obtainable in 128K and 4K context size choices
- Phi-3-small is a 7B parameter mannequin, obtainable in 128K and 4K context size choices
- Phi-3-medium is a 14B parameter mannequin, obtainable in 128K and 4K context size choices
Phi-3-vision is the primary multimodal mannequin within the household, and might generate insights from charts and diagrams. “Phi-3-vision builds on the language capabilities of the Phi-3-mini, persevering with to pack robust language and picture reasoning high quality in a small mannequin,” Bilenko wrote.
In line with Microsoft, in comparison with different fashions, Phi-3 performs nicely. For instance, Phi-3-small beats GPT-3.5T throughout quite a lot of language, reasoning, coding, and math benchmarks, whereas Phi-3-medium beats out Gemini 1.0 Professional. Moreover, Phi-3-vision outperforms Claude-3 Haiku and Gemini 1.0 Professional V generally visible reasoning duties, OCR, desk, and chart understanding duties.
All the Phi-3 fashions are at the moment obtainable on Azure AI and Hugging Face.