one year on
Microsoft unveils Phi-3-mini, a 3.8B model that runs on phones and outperforms models twice its size across key benchmarks
The small language model, trained on curated data, expands the selection of high-quality small models for customers.
Microsoft today released Phi-3-mini, a 3.8-billion-parameter small language model that the company says outperforms models twice its size across a variety of language, reasoning, coding, and math benchmarks. The model is available on Azure AI Studio, Hugging Face, and Ollama, and is optimized to run on-device including mobile hardware via ONNX Runtime and Windows DirectML.
Building on prior work with Phi models (“Textbooks Are All You Need”), Phi-3 models are trained using high-quality data. A 128K-token context variant is also available, a first for a model this size. Microsoft plans to follow with 7B and 14B versions in the coming weeks.
The model’s safety claims are front-loaded: Microsoft says it underwent red-teaming, RLHF, and evaluations across dozens of harm categories, detailed in a technical paper and model cards. One early customer is ITC in India, using Phi-3 for a farmer-facing app called Krishi Mitra. The subtext of today’s announcement is plain: the conversation about cost and accessibility in AI is shifting from ‘how big can we build’ to ‘how small can we make it while still being useful.’
The record
One year later — open only if you can handle spoilers
Phi-3-mini became a reference point for on-device AI, with subsequent models like Apple's OpenELM and Google's Gemma 2 citing similar data-curation strategies. The model's phone-runnable claim proved durable, though adoption was limited by the lack of a permissive license.