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. Microsoft says this release expands the selection of high-quality models for customers.
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.
The Weekly Replay · free by email
This week, one year ago — every Sunday.
One email each Sunday: the week's replayed AI news, with the one-year-later annotations included. Written like it's breaking — dated like it isn't.
Free · double opt-in · unsubscribe anytime · privacy