Local AI hardware, priced today

← swipe the table sideways →

Click a column header to sort. "Usable for model" matters more than raw memory: macOS caps GPU allocation at roughly 75% of unified memory by default (raise it with sudo sysctl iogpu.wired_limit_mb), and Strix Halo is hard-capped at 96GB on Windows but reaches about 120GB on Linux through GTT. Prices are US list or street as of 20 Sep 2026; "~" marks a figure assembled from a verified base price plus a verified upgrade cost rather than read directly off a configurator.

Both tables below are measured numbers from published benchmark runs with stated builds — not vendor claims, not estimates. Prefill and decode are reported separately because they behave completely differently.

gpt-oss 20B · MXFP4 · llama.cpp · pp2048 / tg128

← swipe the table sideways →

* Strix Halo figures are pp512 rather than pp2048 and therefore understate its peak prefill — the two are not strictly comparable. Flagged rather than silently mixed in.

gpt-oss 120B · MXFP4 · the model that needs the memory

← swipe the table sideways →

This table is the argument for unified memory in one screen. A $4,117 RTX 5090 manages 8–10 tok/s on gpt-oss 120B because the model does not fit in 32GB and the experts have to spill to system RAM. A $2,999 Strix Halo box does 33–35 because it does fit. Capacity beats bandwidth the moment the model stops fitting.

Mixture-of-experts vs dense, paired on identical hardware

← swipe the table sideways →

Note the pattern in the last column: the MoE advantage is largest on the most bandwidth-starved machines (8.3x on DGX Spark, 7.5x on Strix Halo) and smallest on the GPU (3.7x on a 5090), because the GPU was never bandwidth-starved to begin with. If you are buying a unified-memory box, you are buying it to run MoE models.

* The DGX Spark row is the largest gap in this table and also the weakest-sourced: the published run reports the figures but states no llama.cpp build hash or test date, unlike every other row. The numbers stay because four independent machines show the same pattern — but treat that specific 8.3x as the least firm number on this page.

What actually fits, at Q4

← swipe the table sideways →

Budget weights plus KV cache, not weights alone — an 8B model at 32K context needs about 4.5GB of KV cache at FP16 on top of its weights, and at long context the cache can exceed the model. Quantizing the cache to q8_0 roughly halves it with little quality cost.

These are real gaps, not oversights. Several pages will happily give you numbers for all of them. Those numbers are invented.

Prices come from the vendor's own store page wherever one exists, and from a named retailer where it does not. Specs come from the manufacturer. Benchmarks come from published runs that state their model and quantization, and in almost every case their build hash — primarily the llama.cpp and MLX discussion threads, plus independent reviewers. Vendor performance claims are excluded from the tables and noted separately where they matter. A cell reads "not verified" when no primary source could be found; nothing is interpolated, and no number is carried over from a machine it wasn't measured on. The single row sourced from a run without a stated build is marked on the table rather than quietly mixed in.

Content-farm aggregators were deliberately excluded. A large number of sites publish hardware tables that recycle each other's figures with no methodology, and several carry numbers that are physically impossible. If a figure here looks surprising, the source is linked below — check it.

Sources