GPT-OSS 20B

GPT-OSS 20B loads in 13.8 GB at MXFP4 and is comfortable from 32 GB of memory, Macs included from the MacBook Air M5 24GB. Below: the full memory math, the cheapest card that runs it well, and the fastest machine we track.

PARAMETERS
21B (3.6B active)
FORMAT
MXFP4
MIN MEMORY
24 GB
BEST FOR
Chat, Coding, Reasoning

Memory math

Weights (MXFP4)
13.8 GB
+ KV cache (16k)
~4.0 GB
Total at 16k
~17.8 GB
Comfortable from
32 GB

KV = fp16 estimate (q8_0 cache roughly halves it). "Comfortable" = weights + KV within the engine's tiered budget (~70-85% of memory).

Hardware snapshot

Cheapest GPU
~62 tok/s est. — ~$550 used
Fastest
~122 tok/s est.
Runs on a Mac?
~25 tok/s est.

Go deeper

More GPT-OSS models

Frequently asked questions

How much memory does GPT-OSS 20B need?

13.8 GB for the MXFP4 weights, plus ~4.0 GB of KV-cache at 16k context — about 17.8 GB total. Comfortable from 32 GB of VRAM or unified memory.

Does GPT-OSS 20B run on a Mac?

Yes — from the MacBook Air M5 24GB (~25 tok/s est.). Unified memory means the RAM budget is the only limit.

What is the cheapest GPU for GPT-OSS 20B?

The AMD Radeon RX 7900 XT is the cheapest tracked card that runs GPT-OSS 20B comfortably — ~62 tok/s est. at ~$550 used.

Cite this page

ModelFit: GPT-OSS 20B — specs, memory math and hardware verdicts.
https://modelfit.io/models/gpt-oss-20b/ (dataset updated 2026-09-03, CC BY 4.0).