What hardware do you need to run gpt-oss-20b?

gpt-oss-20b at MXFP4 is 12 GB of weights, with a context ceiling of 128k tokens. Punchy for 12 GB, with adjustable reasoning effort. High effort is slow; low effort is shallow. On the cheapest machine that runs it, a Mac mini M6, 24GB, it generates 18 tok/s at 32k of context, worked out from memory bandwidth.

How good is gpt-oss-20b, really?

On the Artificial Analysis Intelligence Index v4.3 it scores 9 (high reasoning effort; 10 at low), which puts it in the Below every hosted tier band. Fine for simple, well-specified tasks. Noticeably less capable than anything the big labs sell today. See the whole table.

• Summarising — good
• Translation — usable
• Everyday coding — good
• Reasoning & maths — usable
• Agentic work — usable

Head to head with its neighbours on the leaderboard: vs Qwen3-Coder Next above it, vs Gemma 4 E4B below. One more model needs much the same memory at 32k of context: vs Ministral 3 14B.

What gpt-oss-20b costs either way

Renting the same model costs $0.018 per million input tokens and $0.09 per million output (OpenRouter, cheapest active endpoint, checked 2026-10-29). Buying a machine only beats that if you use it hard enough, for long enough, that the hardware price divides down below the rental bill. What a million tokens costs each way puts the two prices side by side.

Machines that run gpt-oss-20b

Every machine here holds it to 128k, this model's own ceiling, so the choice between them is speed and price rather than how much you can keep in the window.

The table is the cheapest machine in each family, not the only one that runs it: 50 of the 56 machines on this site hold it at 32k. Within a family only memory decides, so that is every Mac Studio and Strix Halo box at any size, every NVIDIA card with 16GB or more, every Mac mini and MacBook Pro with 24GB or more, the Radeon AI PRO R9700, 32GB and the DGX Spark, 128GB. One table sets every machine against the others: price, the memory its GPU can use, what it holds and how long it takes to pay for itself.

One machine per family, cheapest first. Speeds are measured where a public benchmark exists and estimated from memory bandwidth otherwise; the calculator says which for any configuration. The longest context is the longest setting the calculator offers that the machine still holds this model at, cache included, and no machine is shown taking it past its own 128k limit. Each one opens the calculator on that machine at that length.

For scale, the calculator starts from 80 tok/s for a hosted API and times a local machine against it. 2 of the 7 above reach it, and the slowest is 16 tok/s.

The cheapest machine that runs it is sold with 24 GB. There is a page on what 24 GB runs, machine by machine: it names the 7 machines here sold at that size, and the 28 current models the roomiest of them holds at 32k, this one among them.

It fits in less memory than that: the smallest size here that runs it at 32k is 16 GB, on the GeForce RTX 4080, 16GB at $1,199, card only, which hands a model 15 GB and is no longer sold.

Run gpt-oss-20b at 128k on the Mac mini M6, 24GB

How much memory does gpt-oss-20b need at 128k?

Every figure above is priced at 32k of context. At 128k, this model's own ceiling, gpt-oss-20b needs 15.3 GB rather than 12.9 GB. The weights are the same either way; the key-value cache goes from 0.8 GB to 3.2 GB.

The window costs nothing in hardware: the same 33 of the 37 machines here still sold at a published price hold gpt-oss-20b at both lengths, and the cheapest of them is the Mac mini M6, 24GB at $1,099.

On the Mac mini M6, 24GB it runs 18 tok/s estimated at 32k and 9.6 tok/s estimated at 128k. Every token generated reads the whole cache back, so a longer window is slower as well as bigger.

What has been measured running gpt-oss-20b

Three of the speeds in the table above were measured, and there are more measurements than that table shows. 11 machines here have a published speed for gpt-oss-20b. The table above has three of them; of the other eight, five are no longer sold and three are still sold but not the cheapest of their family. They are below, as published and again at 32k of context, which is what every other speed on this page is taken at.

A published figure is what its source reports, at the window the model was run at. The 32k figure is that measurement read at 32k of context: every token generated reads the weights and the whole key-value cache, so the longer the window the slower it runs. A discontinued machine's price is the one it launched at. Where two memory tiers of one machine carry the same figure, the speed is set by the bandwidth of the chip they share rather than by the amount of memory. Each published figure links the source that took it.

The specifics

Parameters

20.9B, of which 3.6B are active per token

Quantisation

MXFP4

Weights on disk

12 GB

KV cache

0.8 GB at 32k context — Half the layers use a 128-token sliding window; KV cache is tiny. How weights and cache add up.

Maximum context

128k tokens

Licence

Apache 2.0

Sources

ggml-org/gpt-oss-20b-GGUF, openai/gpt-oss-20b