Small Models, Big Impact

Learning Objectives Understand the small model revolution and why it matters Know when small models are a better choice than frontier models See the practical implications of running AI locally The Counterintuitive Trend While the headlines focus on ever-larger AI models costing billions to train, a quieter revolution is happening: small models are getting shockingly good. In 2024, a model with 7 billion parameters was barely useful for serious work. In 2026, models with 3 to 8 billion parameters can handle most everyday tasks — writing, summarizing, coding, answering questions — at a quality level that would have been considered state-of-the-art just two years ago. This matters because small models can run on a laptop, a phone, or a cheap server. No cloud subscription. No internet connection. No data leaving your device. The Key Small Models Microsoft Phi Series Microsoft's Phi models are designed to punch above their weight — small models trained on high-quality data that compete with much larger models on many tasks. Phi-4 — 14 billion parameters, strong at reasoning and coding Available via Azure, Hugging Face, and increasingly in Microsoft products Designed for scenarios where cost, speed, or privacy matter more than maximum capability Google Gemma Google's open-source small model family: Gemma 2 — available in 2 billion and 9 billion parameter versions Strong performance relative to size, especially for structured tasks Runs on consumer hardware, integrates with Google's AI ecosystem Meta Llama (Small Variants) Meta's Llama 4 family includes models from 8 billion parameters up…