Recommended setups

Four machines that clear 20 tok/s on our estimate. Each card shows the largest Q4 model that fits without dropping below that floor.

Mac mini M4

Apple desktop

M4, 16 GB unified memory

Runs
gpt-oss-20b 10.5 GB Q4
Speed
~25-35 tok/s
Price
$599 new

The cheapest way to clear 20 tok/s. 16 GB is a tight fit, and you supply the monitor.

See at Apple checked Nov 17, 2026

MacBook Pro M4 Pro

Apple laptop

M4 Pro, 24 GB unified memory

Runs
Qwen3-30B-A3B 15.5 GB Q4
Speed
~70-95 tok/s
Price
$1,999 new

The portable daily driver. A tight 24 GB fit, but twice the memory bandwidth of the mini.

See at Apple checked Nov 17, 2026

RTX 4060 Ti 16 GB

CUDA desktop

16 GB VRAM, 32 GB system memory

Runs
Qwen3-30B-A3B 15.5 GB Q4
Speed
~60-80 tok/s
Price
$1,200 build

The value CUDA desktop. The pick leaves little headroom, so context is tight.

See at NVIDIA checked Nov 17, 2026

RTX 4090 24 GB

CUDA desktop

24 GB VRAM, 32 GB system memory

Runs
Qwen3.5-35B-A3B 17.5 GB Q4
Speed
~210-285 tok/s
Price
$2,000+ build

The fast option. It is also the one that needs the most power and space.

See at NVIDIA checked Nov 17, 2026

Prices are starting points, not quotes. We do not take affiliate money for these. Check what your machine runs.