The “Phi-3 Technical Report: A Highly Capable Language Model Locally on Your Phone” was submitted to arXiv on April 22, 2024 by Microsoft Research. It introduced the Phi-3 family, whose smallest member, Phi-3-mini, has just 3.8 billion parameters yet reaches 69 percent on the MMLU benchmark, performance the authors compare to GPT-3.5 and Mixtral 8x7B despite being a fraction of the size. The family also includes Phi-3-small (7B), Phi-3-medium (14B), and later mixture-of-experts and vision variants.
The reported secret is data quality rather than scale. Phi-3 continues the “Textbooks Are All You Need” line of work, training on a heavily filtered and curated mix of web data and synthetic, textbook-style content designed to be educational and reasoning-rich. The result is small enough to be quantized and deployed directly on a phone while still being broadly capable.
Phi-3 is a leading example of the small-language-model trend: instead of chasing ever-larger models, optimize the training data so a compact model punches above its weight. For organizations, capable models that run locally on phones and laptops mean lower cost, lower latency, and the ability to keep data on-device rather than sending it to a cloud API.