Find the best hardware for running AI locally

Not just can this model fit, but: how fast will it actually go, what will it cost you to own after tax and electricity, and is there a better option — including not buying anything at all.

Start here
Hardware configurations
21
10 products
Open-weight models
27
96 quantized variants
Measured benchmarks
15
every one traced to a source
Estimated pairings
1,163
labelled, never precise
The rule this site is built on: never hide uncertainty. Measured numbers are precise. Estimates are shown as ranges and marked Estimated. Assumptions are marked Assumption and you can change them. Manufacturer figures are marked Official spec because a theoretical bandwidth is not a measurement. Where four sources disagree, we show the spread instead of an average.
Recent measured benchmarksAll benchmarks →
HardwareModelQuantContextDecodePrefillTypeDate
Framework Desktop (Ryzen AI Max+ 395, 128 GB)gpt-oss-120bMXFP44K34.0 t/sMeasured (community)11 May 2026
GMKtec EVO-X2 (Ryzen AI Max+ 395, 128 GB)gpt-oss-120bMXFP44K32.0 t/sMeasured (community)20 Apr 2026
Framework Desktop (Ryzen AI Max+ 395, 128 GB)gpt-oss-120bMXFP44K31.0 t/sMeasured (community)2 Apr 2026
Framework Desktop (Ryzen AI Max+ 395, 128 GB)gpt-oss-120bMXFP44K55.0 t/sMeasured (community)15 Mar 2026
Framework Desktop (Ryzen AI Max+ 395, 128 GB)gpt-oss-120bMXFP44K30.0 t/sMeasured (community)10 Feb 2026
Framework Desktop (Ryzen AI Max+ 395, 128 GB)Qwen3.6 27BQ4_K_M4K11.5 t/sMeasured (community)10 Feb 2026
Framework Desktop (Ryzen AI Max+ 395, 128 GB)Qwen3-Coder 30B-A3BQ4_K_M4K52.0 t/sMeasured (community)10 Feb 2026
RTX 5090 workstation (1x 32 GB)Qwen3 30B-A3BQ4_K_M0352.1 t/sMeasured (community)30 Nov 2025
Popular comparisonsAll →