Running Phi models on a NAS

Which Phi model should you run on a NAS?

Phi-4-mini is small and fast, around 2.5GB at 4-bit, with a long context window for documents. It suits 16GB systems and quick tasks. Phi-4 14B, about 9.1GB at 4-bit, is noted for reasoning and maths well above its size, making it a good fit for 32GB systems that run models on the CPU (Estimated).

Why do small models suit a NAS?

Most NAS systems run models on the processor rather than a graphics card. Small models answer quickly on that hardware, so a well-trained 4B or 14B model often feels better day to day than a larger model that answers slowly.

Does Phi need a GPU?

No. Both models run comfortably on modern AI NAS processors.

What can Phi do on a NAS?

Phi-4 handles summaries, drafting, maths and structured answers well, and Phi-4-mini is quick enough to sit behind a home automation or document search tool without slowing other services. Neither model reads images, so pair them with Gemma 3 if you want to ask about photos or scanned pages. For long documents, Phi-4-mini's larger context window is the better choice of the two.

Common questions

Which quantizations are published?

Ollama publishes Q4, Q8 and full precision for both Phi-4 models.

Is Phi free to use?

Phi-4 is released under the MIT licence.