Phi 5.
Microsoft's small-language-model family for on-device and edge reasoning at 14B parameters.
What it is.
Microsoft Research's small-language-model line. Phi 5 ships at 14B dense parameters with frontier-comparable reasoning on the MMLU and HumanEval benchmarks Microsoft published in 2025. Distilled from larger teacher models with curated synthetic data. Runs on a Copilot+ PC NPU at interactive latency. Distributed under MIT licence.
Where it fits.
Anywhere the buyer needs reasoning without a cloud round-trip. Copilot+ PCs ship Phi 5 Mini as the default on-device model. Retail edge deployments, vehicle infotainment, and call-centre desk software now treat Phi 5 as the cost-anchor for local inference.
- Strongest open-source performance per parameter
- MIT licence removes commercial friction
- Runs on NPU and consumer GPU at interactive latency
- Behind frontier-closed on long-tail world knowledge
- Smaller context window than Qwen or Llama 4
Frequently asked.
Why use Phi 5 over Llama 4?
Smaller footprint and stronger reasoning per parameter. Phi 5 Mini fits in 8GB of NPU memory, which Llama 4 cannot match without aggressive quantisation.
Is Phi 5 available on Azure?
Yes. Azure AI Foundry serves Phi 5 as a managed endpoint with enterprise governance. Microsoft also ships the weights to Hugging Face.
What is the training data approach?
Curated synthetic data generated by GPT-4-class teachers, filtered against textbook-quality benchmarks. Microsoft documented the recipe in the Phi-3 technical report and extended it for Phi 5.
