Microsoft Phi-4 Mini: GPT-4 Level Performance in 3.8B Parameter Model

Microsoft Research released Phi-4 Mini (3.8B parameters), demonstrating that careful training data curation enables small models to achieve quality comparable to GPT-4 on reasoning-heavy tasks. Phi-4 Mini achieves 72.3% on MMLU and 75.4% on GPQA, exceeding much larger models. The model runs on consumer hardware: 4GB RAM for CPU inference, runs at 30+ tokens/second on Apple M-series chips. Key applications: edge AI (phones, laptops, IoT without cloud), cost-efficient inference at scale, specialized fine-tuning for domain-specific tasks. Available on Hugging Face under MIT license.