Microsoft
Can the Arc B580 run Phi-4 14B?
We earn from qualifying purchases via affiliate links. Disclosure.
Yes, but it's tight. Phi-4 14B is a 9.1 GB download (Q4_K_M). The Arc B580 gives the GPU 12 GB, leaving about 2.9 GB, so keep the context short (our estimate).
Does it fit?
| Phi-4 14B memory | 9.1 GB at Q4_K_M (Ollama: phi4 tags) |
|---|---|
| Arc B580 GPU memory | 12 GB of VRAM (product listing). |
| Left over | About 2.9 GB (our arithmetic) |
| Verdict | Tight Under 4 GB to spare (our estimate bands) |
Recommended quantization and context
Use the Q4_K_M build that Ollama's tag downloads; the memory figure above is for that build. The model supports up to 16K of context, but longer context uses more memory. Ollama starts at a 4096-token context by default (Ollama FAQ).
With only about 2.9 GB spare, keep the default context. If ollama ps shows part of the model on the CPU, shorten the context or pick a smaller model.
Run it
ollama run phi4:14b Then check it's on the GPU:
ollama ps The PROCESSOR column should read 100% GPU. Driver setup for Intel cards on Windows, Linux and macOS, plus common fixes, is in the install and run section.
Where to buy the Arc B580
ASRock Challenger Intel Arc B580 12GB
$329.99 Newegg · checked Oct 27, 2026, about 12:35 PM CT
Sold by Newegg, seen on the product page
Check price at Newegg (opens in a new tab)FAQ
Can the Arc B580 run Phi-4 14B?
Yes, but tight: Phi-4 14B needs 9.1 GB (Q4_K_M, Ollama library) and the Arc B580 gives the GPU 12 GB, leaving about 2.9 GB. Keep the 4096-token default context. The fit bands are our estimate.
How much VRAM does Phi-4 14B need?
9.1 GB for the Q4_K_M build in the Ollama library, plus roughly 1–4 GB for context and runtime (our estimate). See the 12 GB page for other models that fit.
What command runs Phi-4 14B?
ollama run phi4:14b. Then run "ollama ps" and check the PROCESSOR column reads 100% GPU.
How fast is Phi-4 14B on the Arc B580?
We don't list a speed because we have no sourced measurement for this pairing. Speed depends on memory bandwidth, quantization and context length.