Cheapest Mac to run GPT-OSS 20B
21B parameters (3.6B active) · quality index 33 · coding 30 · every Apple Silicon Mac ever sold, compared. It is a mixture-of-experts model: all 21B parameters must sit in memory, but only 3.6B compute per token, which is why it is faster than dense models of similar size.
The cheapest Mac that runs GPT-OSS 20B comfortably is a used Mac mini M4 10c/10c 24GB (2024) at about $615 EST. on the used market — running Q6_K quantization at roughly 21 tok/s with up to 15K tokens of context.
Every Mac that runs it, by used price
| Machine | Used price | Runs at | Est. speed | Max context |
|---|---|---|---|---|
| Mac mini M4 10c/10c 24GB (2024) | $615 EST. | Q6_K | 21 tok/s | 15K tokens |
| Mac mini M2 8c/10c 24GB (2023) | $624 EST. | Q6_K | 18 tok/s | 15K tokens |
| Mac mini M4 10c/10c 32GB (2024) | $769 EST. | Q6_K | 21 tok/s | 54K tokens |
| Mac mini M2 Pro 10c/16c 32GB (2023) | $1,062 EST. | Q6_K | 36 tok/s | 54K tokens |
| Mac mini M4 Pro 12c/16c 24GB (2024) | $1,077 EST. | Q6_K | 49 tok/s | 15K tokens |
| Mac Studio M1 Max 10c/24c 32GB (2022) | $1,229 EST. | Q6_K | 71 tok/s | 54K tokens |
| Mac mini M4 Pro 14c/20c 24GB (2024) | $1,231 EST. | Q6_K | 49 tok/s | 15K tokens |
| Mac Studio M2 Max 12c/30c 32GB (2023) | $1,249 EST. | Q6_K | 71 tok/s | 54K tokens |
| Mac mini M2 Pro 12c/19c 32GB (2023) | $1,249 EST. | Q6_K | 36 tok/s | 54K tokens |
| Mac Studio M1 Max 10c/32c 32GB (2022) | $1,352 EST. | Q6_K | 71 tok/s | 54K tokens |
See all 63 machines, other price bases, and live currency conversion →
Run it
ollama run gpt-oss:20b pulls the default (≈Q4) build once you have Ollama installed.
Or skip the hardware
Cloud APIs serve GPT-OSS 20B at about $0.2 per million output tokens. The main site's break-even solver computes the daily usage where owning a Mac becomes cheaper than renting.
Open the interactive guide — speed simulator, TCO, all 75 machines →