Mini instruct

A lightweight 3.8-billion parameter open model optimized for high-performance reasoning and precise instruction following on constrained hardware.

Microsoft's Phi-4-mini-instruct delivers state-of-the-art reasoning capabilities in a highly compact 3.8B parameter architecture. Built on a dense transformer foundation and trained on high-quality, reasoning-dense synthetic and curated web data, this model supports an expansive 128K token context window. It excels in memory-constrained and latency-bound environments (such as local edge devices and mobile hardware) by utilizing supervised fine-tuning and direct preference optimization to match the instruction-adherence quality of much larger systems.

https://huggingface.co/microsoft/Phi-4-mini-instruct