BuildsGPUsModelsGuidesMethodologyNewsletter

Can the Arc B580 run phi-4?

Yes. The Q4_K_M GGUF of phi-4 is 9.05 GB (source). With our 1.5 GB allowance for runtime and short-context cache it needs about 10.6 GB, leaving 1.4 GB of the card's 12 GB free. Expected speed: ~38 tok/s estimate. Lowest tracked price: low $330, avg $446 (new) source, 2026-10-16, which works out to $9 per tok/s (#1 of 16 priced GPUs that fit this model).

Last updated Oct 16, 2026. Speed is a bandwidth-based estimate (456 GB/s, TechPowerUp), not a benchmark. Long contexts need more memory than our allowance.

Tight fit: under 3 GB spare means long prompts or a bigger context window may not fit. Consider a smaller quant or a card with more VRAM.

Next size up: gpt-oss-20b (11.67 GB at Q4_K_M) needs ~13.2 GB and does not fit on one Intel Arc B580. See what runs gpt-oss-20b.

Every priced GPU that runs phi-4, by cost per token/sec

GPUVRAMSpeed estimateLow price$ per tok/s
Intel Arc B58012 GB~38 tok/s$330 (new)$9
AMD Radeon RX 907016 GB~54 tok/s$570 (new)$11
AMD Radeon RX 7800 XT16 GB~52 tok/s$599 (new)$12
NVIDIA GeForce RTX 507012 GB~56 tok/s$800 (new)$14
NVIDIA GeForce RTX 5070 Ti16 GB~75 tok/s$1,120 (new)$15
NVIDIA GeForce RTX 3060 12 GB12 GB~30 tok/s$455 (used)$15
AMD Radeon RX 9070 XT16 GB~54 tok/s$810 (new)$15
AMD Radeon RX 7900 XTX24 GB~81 tok/s$1,350 (new)$17
NVIDIA GeForce RTX 508016 GB~81 tok/s$1,600 (new)$20
NVIDIA GeForce RTX 5060 Ti 16 GB16 GB~38 tok/s$780 (new)$21
NVIDIA GeForce RTX 4070 SUPER12 GB~42 tok/s$966 (used)$23
AMD Radeon RX 9060 XT 16 GB16 GB~27 tok/s$620 (new)$23
NVIDIA GeForce RTX 309024 GB~79 tok/s$1,889 (used)$24
NVIDIA GeForce RTX 409024 GB~85 tok/s$3,200 (used)$38
NVIDIA GeForce RTX 509032 GB~150 tok/s$6,500 (new)$43
NVIDIA RTX A600048 GB~64 tok/s$6,170 (new)$96

Prices: picksbycard read of Newegg listings, 2026-10-16. Prices move daily.

Buy on Amazon As an Amazon Associate I earn from qualifying purchases.

Related