gpt-oss-20b

OpenAI

Free

Get

MLX Score4.7

PriceFreeno paid plan

RunsOn your Mac

DeveloperOpenAI

LicenceOpenApache-2.0

CategoryModels

Mac14.0+macOS or later

gpt-oss-20b is the smaller of OpenAI's two open-weight models, made for lower latency and local use. It has 21 billion parameters with 3.6 billion active per token, and you can set its reasoning effort to low, medium or high.

It ships in a native 4-bit format (MXFP4); the MLX build is about 12 GB. On mlx.app's estimate it needs a Mac with 24 GB of memory.

Open weights under Apache 2.0. Run locally, nothing leaves your Mac.

How to Get It

Run it on your MacSee how

WebsiteVisit

Official downloads only. MLX never re-hosts an app, and never runs a command for you.

MLX Score

4.7out of 5

Hugging Face downloads and likesahead of 93%

Hugging Face: 5,939,725 downloads (30 days), 5,263 likes · read 2026-11-02

Ranked against the open models listings, from real public data only. How the score works.

Information

Developer: OpenAI
Category: Open models
Price: Free (Open weights, free to download.)
Your data: Runs on your Mac
Licence: Apache-2.0
Mac: macOS 14.0 or later
iPhone and iPad: —
Needs: 24 GB memory
Links: Website · Hugging Face
Facts checked: 2026-11-01

Sources

huggingface.co/openai/gpt-oss-20b · Apache-2.0, 21B parameters with 3.6B active, local use, configurable reasoning effort · checked 2026-11-01huggingface.co/mlx-community/gpt-oss-20b-MXFP4-Q8 · MLX build MXFP4-Q8 (about 12 GB) · checked 2026-11-01ml-explore.github.io/mlx/build/html/install.html · macOS 14.0 minimum (MLX needs Apple silicon and macOS 14.0+) · checked 2026-11-01mlx.app/run/gpt-oss-20b · memory: smallest Mac (24 GB) where the 4-bit MLX build fits with room for context, from lib/compare (75% of unified memory usable) · checked 2026-11-01Icon from Hugging Face

See All