Phi-3

Definition

Phi-3 is a family of small language models from Microsoft Research, released in 2024, designed to deliver strong reasoning and language performance at sizes small enough to run efficiently on phones and other edge devices.

Overview

Phi-3 continued Microsoft Research's 'Phi' line of small language models, built on the premise that careful curation of high-quality, textbook-like training data can let a much smaller model punch above its parameter count on reasoning, coding, and language benchmarks, rather than relying purely on scaling up size. The family included multiple sizes — Phi-3-mini, Phi-3-small, and Phi-3-medium — with Phi-3-mini in particular notable for running on a modern smartphone while still delivering performance comparable to much larger models on many tasks. Microsoft released Phi-3 with open weights, making it available for download, self-hosting, and fine-tuning through platforms such as Hugging Face and Azure AI, and it is positioned as a practical option for on-device AI, latency-sensitive applications, and cost-constrained deployments where a large frontier model would be unnecessary or too expensive. Phi-3 sits within Microsoft's broader AI strategy alongside its partnership with OpenAI and Azure's hosting of other third-party models, giving Microsoft customers a spectrum of options from tiny efficient foundation models like Phi-3 up to frontier-scale models accessed via Azure OpenAI.

Key Features

• Small language models trained on curated, high-quality data for efficiency
• Multiple sizes: Phi-3-mini, Phi-3-small, and Phi-3-medium
• Phi-3-mini capable of running on modern smartphones
• Openly released weights for self-hosting and fine-tuning
• Available via Hugging Face and Azure AI
• Strong reasoning and coding performance relative to model size

Use Cases

Frequently Asked Questions

Related in AI Models