AI modelsOpenAI

gpt-oss-20b

Decision matrix

gpt-oss-20b model decision matrix
Overview

gpt-oss-20b is OpenAI's smaller open-weight model (21B total, 3.6B active parameters, mixture-of-experts), Apache 2.0 licensed, with a 128K-token context window, running on 16GB of VRAM or unified memory, so it fits on a capable consumer GPU or Apple Silicon Mac. Free to download and use commercially; the realistic entry point into the gpt-oss family for local use.

Other good fits
  • A first test of the gpt-oss family before deciding whether a task needs the larger gpt-oss-120b.
Pros
Read pros

gpt-oss-20b is OpenAI's smaller open-weight model: 21 billion total parameters in a mixture-of-experts architecture, with 3.6 billion active per token across 32 experts, released under the Apache 2.0 licence. Its hardware requirement, 16GB of VRAM or unified memory, puts it within reach of a capable consumer GPU or an Apple Silicon Mac, unlike the larger gpt-oss-120b's 80GB-class requirement.

Pick it when…

  • You want to try an OpenAI-family open-weight model locally without needing workstation-class GPU hardware.
  • You're testing whether a smaller local model is good enough for a task before considering the larger gpt-oss-120b or a hosted API.
  • You want commercial-use rights on a model you run and control yourself.
Cons
Read cons

A 20B-class model, even a capable one, won't match a much larger model's reasoning depth on genuinely hard tasks; test on your actual workload before committing to it for anything demanding. Like gpt-oss-120b, the training data and pipeline aren't published, so you can run and fine-tune it but can't audit exactly how it was originally trained.

Pick something else when…

  • The task needs more capability than a 20B model reliably delivers. gpt-oss-120b is the larger sibling, if your hardware can handle it.
  • You'd rather not manage local inference at all. A hosted API, GPT-mini or similar, removes the hardware requirement entirely.
  • You want an easy way to try it without command-line setup. LM Studio is built for exactly that.

Avoid if

  • You'd rather not manage local inference at all; a hosted API removes that setup entirely, at a per-token cost.
Context

gpt-oss-20b has a 131,072-token (128K) context window, the same as gpt-oss-120b, with three selectable reasoning-effort levels. The Apache 2.0 licence covers the released weights and inference code; training data and pipeline details aren't published.

Requirements

Not documented yet.

Licence & cost

The model weights are free to download and use commercially under Apache 2.0; the real cost is hardware, specifically at least 16GB of VRAM or unified memory, achievable on many higher-end consumer GPUs or Apple Silicon Macs with enough RAM. There's no subscription or per-token fee for local use; the ongoing cost is whatever electricity and hardware you're already running.

First task

Documentation-based suggestion — not run here.

Install a local runtime (Ollama or LM Studio) and run gpt-oss-20b on a made-up drafting or summarisation task, then compare its output against the same task run on a hosted model you already have access to, noting both quality and how the local model's speed felt on your specific hardware.

Done when you've compared local gpt-oss-20b output against a hosted model on the same task and can judge whether local, self-hosted use is good enough for that kind of work.

Links & evidence
Documentation-based · not hands-on · Sep 23, 2026

Rests on the gpt-oss GitHub repository and OpenAI's cookbook guide for running it locally with Ollama, both read on 2026-09-23. No model was downloaded or run for this review.

  • gpt-oss GitHub repository · Sep 23, 2026 — 21B total/3.6B active parameters, 32 experts, 128K context window, Apache 2.0 license.
  • Running gpt-oss locally with Ollama · Sep 23, 2026 — Hardware guidance: gpt-oss-20b best with ≥16GB VRAM/unified memory.
  • gpt-oss-20b on Hugging Face · Sep 23, 2026 — Model card and weight distribution.
Further review notes & alternatives

Alternatives and trade-offs

  • gpt-oss-120b wins when your hardware can reach 80GB-class GPU memory and the task needs more capability.
  • Gemma 4 wins as an alternative open-weight family at similar hardware tiers, worth comparing on your specific task.
  • LM Studio wins as the easiest way to actually run gpt-oss-20b locally without command-line setup.

Learn more

Catalogue facts

Run an open-weight model locally for private tool-use tasks. Use its 128K context for substantial prompts and documents. Choose it when your machine has about 16 GB of memory.

OpenAI · free · self-host

  • laptop inference
  • low vram
  • moe architecture
  • 4-bit quantization
  • open weight
  • apache licence
  • offline
  • write-code
  • automate
  • open-source
  • talk-think
  • draft-docs

Browse all tools