AI modelsOpenAI

gpt-oss-120b

Decision matrix

gpt-oss-120b model decision matrix
Overview

gpt-oss-120b is OpenAI's larger open-weight model (117B total, 5.1B active parameters, mixture-of-experts), Apache 2.0 licensed, with a 128K-token context window, runnable on a single 80GB GPU. Free to download and use commercially; pick it when you want an OpenAI-family model you can self-host, with real hardware requirements to plan around.

Other good fits
  • Someone with access to a single 80GB-class GPU who wants more capability than the smaller gpt-oss-20b offers.
Pros
Read pros

gpt-oss-120b is OpenAI's larger open-weight model release: 117 billion total parameters in a mixture-of-experts architecture, with only 5.1 billion active per token across 128 experts, released under the permissive Apache 2.0 licence. Unlike OpenAI's closed API models, you can download the weights, run them yourself and use them commercially without a per-token fee to OpenAI.

Pick it when…

  • You want an OpenAI-family model you can self-host, in your own infrastructure or a private cloud, rather than calling OpenAI's API.
  • You have access to a single 80GB-class GPU (or equivalent unified memory) and want more capability than the smaller gpt-oss-20b.
  • You want to fine-tune or adapt the model for a specific task, which the open weights and Apache 2.0 licence allow.
Cons
Read cons

A single 80GB-class GPU is a real hardware requirement, well beyond a typical laptop or even many desktop GPUs, so self-hosting this specific model needs either serious owned hardware or a cloud GPU budget. As with any open-weight release, the training data and pipeline aren't published, so you can't audit or reproduce exactly how it was trained, only what it outputs.

Pick something else when…

  • You don't have 80GB-class GPU access and don't want to rent cloud GPU time. gpt-oss-20b fits in 16GB and is a more realistic local option.
  • You want the model without managing hardware at all. A hosted provider serving gpt-oss-120b through an API removes that, at a per-token cost.
  • You need OpenAI's most capable model regardless of self-hosting. GPT-5 through the API remains ahead of the open-weight line.

Avoid if

  • You need the absolute latest capability; OpenAI's closed flagship models (GPT-5) remain ahead of the open-weight gpt-oss line.
Context

gpt-oss-120b has a 131,072-token (128K) context window and ships with three selectable reasoning-effort levels, letting you trade response depth against speed and cost per request. The Apache 2.0 licence covers the released weights and inference code; the training data and full training pipeline aren't published, so you can run and fine-tune the model but not reproduce its training from scratch.

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. It's documented to run best on 60GB or more of VRAM or unified memory, in practice a single 80GB-class GPU (or a multi-GPU/workstation setup). Running it on rented cloud GPU time instead of owned hardware shifts the cost to an hourly or per-second cloud rate; several hosted providers also serve gpt-oss-120b through their own APIs at a per-token price if you'd rather not manage the hardware at all.

First task

Documentation-based suggestion — not run here.

Run gpt-oss-120b through a local or cloud GPU setup on a made-up long-document summarisation task that uses a meaningful chunk of its 128K context window, then compare the output's accuracy against the source document.

Done when you've checked the summary against a long (fictional) source document and confirmed it didn't lose or invent anything from the parts you specifically checked.

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

Rests on the gpt-oss GitHub repository, its Apache 2.0 licence file and OpenAI's cookbook guide for running it locally with Ollama, all read on 2026-09-23. No model was downloaded or run for this review.

  • gpt-oss GitHub repository · Sep 23, 2026 — 117B total/5.1B active parameters, 128 experts, 128K context window, three reasoning-effort levels.
  • Apache 2.0 license (gpt-oss-120b) · Sep 23, 2026 — Permissive commercial-use license for the released weights and inference code.
  • Running gpt-oss locally with Ollama · Sep 23, 2026 — Hardware guidance: gpt-oss-120b best with ≥60GB VRAM/unified memory.
Further review notes & alternatives

Alternatives and trade-offs

  • gpt-oss-20b wins when your hardware can't reach 80GB-class GPU memory; it fits in 16GB instead.
  • GLM wins as an alternative open-weight-adjacent family to compare capability and licensing against.
  • GPT-5 wins when you need OpenAI's most capable model and are fine calling it through the API instead of self-hosting.

Learn more

Catalogue facts

Run an open-weight reasoning model on substantial local hardware. Use its 128K context for larger local tasks. Choose it when self-hosting matters more than laptop portability.

OpenAI · free · self-host

  • mixture-of-experts
  • 4-bit quantization
  • self-hosted
  • single-gpu
  • high-context
  • open-source licence
  • amd mini-pc
  • token throughput
  • write-code
  • automate
  • open-source
  • talk-think
  • draft-docs
  • data-analytics

Browse all tools