gpt-oss

gpt-oss is OpenAI's family of open-weight models: gpt-oss-120b and gpt-oss-20b, released Aug 5, 2025 under the Apache 2.0 license, which run on one 80 GB GPU or a 16 GB machine. As of Nov 6, 2026 OpenAI has released no newer general open-weight model.

gpt-oss is OpenAI's open-weight model family: gpt-oss-120b and gpt-oss-20b, released on Aug 5, 2025 under the Apache 2.0 license, so anyone can download, run, change and sell products built on them. The 120b model fits on a single 80 GB GPU and the 20b model "can run on edge devices with just 16 GB of memory". As of Nov 6, 2026, no 2026 successor has appeared: OpenAI's later open releases are two safety classifiers built on gpt-oss and small research models.

Quick facts

Released	Aug 5, 2025
License	Apache 2.0, plus a minimal usage policy
gpt-oss-120b	117B parameters, 5.1B active per token; one 80 GB GPU
gpt-oss-20b	21B parameters, 3.6B active; 16 GB of memory
Context	131,072 tokens (128K)
Knowledge cutoff	Jun 1, 2024
Input	Text only
Reasoning effort	Low, medium, high
Ollama downloads	14.7 million, read Nov 6, 2026

What are the gpt-oss models?

Both are mixture-of-experts reasoning models: only a small share of their weights work on each token, which is why the 120b model runs at the speed of a much smaller one. Their expert layers ship in a 4-bit format (MXFP4), which is what lets them fit on one data-center GPU or a well-equipped laptop (Hugging Face). They read and write text only; they do not take images.

They are built for tools and agents. The model card lists built-in web browsing, function calling, Python code execution and Structured Outputs, and you set the reasoning level in the system prompt, for example "Reasoning: high". They must be prompted in OpenAI's harmony format; the model card warns it "will not work correctly otherwise". Common runtimes apply the format for you.

How good are they?

OpenAI's launch post says gpt-oss-120b matches or exceeds o4-mini on Codeforces, MMLU, Humanity's Last Exam and TauBench, and beats it on HealthBench and AIME 2024 and 2025, while gpt-oss-20b matches or beats o3-mini on the same tests. These are OpenAI's own comparisons with its 2025 models; they say nothing about how gpt-oss compares with today's GPT-6 models. No dated 2026 leaderboard result for gpt-oss was read for this page. Google's Antigravity coding tool still offers GPT-OSS-120B but plans to remove it on Nov 18, 2026.

How do you run gpt-oss?

The quickest way on your own machine is Ollama, where the 20b download is 14 GB and the 120b download 65 GB:

Install Ollama from ollama.com.
Run ollama run gpt-oss:20b (or ollama run gpt-oss:120b on a machine with enough memory).
Chat in the terminal, or call http://localhost:11434/api/chat from your own code.

Other ways, from the model card:

LM Studio: lms get openai/gpt-oss-120b, or search for gpt-oss in the app.
Hugging Face Transformers: pip install -U transformers kernels torch, then transformers serve and transformers chat localhost:8000 --model-name-or-path openai/gpt-oss-120b.
vLLM: vllm serve openai/gpt-oss-120b, for serving many users.
llama.cpp, and hosted providers including Azure, AWS, Fireworks, Together AI, Cloudflare and OpenRouter (OpenAI).

OpenAI's own API lists both models on the Responses and Batch endpoints, but gpt-oss-20b shows 0 requests per minute on every tier, so in practice you run it elsewhere. The 120b model page lists it on Responses and Batch.

Safety

Releasing weights means anyone can retrain away the safeguards, so OpenAI tested that case before release: it fine-tuned versions on biology and cybersecurity data to make them stop refusing, and reports that under its Preparedness Framework those versions did not reach high capability levels. Three outside expert groups reviewed the method, and OpenAI ran a $500,000 red-teaming challenge. On Oct 29, 2025 it released gpt-oss-safeguard-120b and gpt-oss-safeguard-20b, open-weight safety classifiers fine-tuned from gpt-oss.

Frequently asked questions

Is gpt-oss free? Yes. OpenAI released gpt-oss-120b and gpt-oss-20b under the Apache 2.0 license, so the weights are free to download, run, modify and use commercially. You pay only for your own hardware or a hosting provider.

What hardware does gpt-oss need? gpt-oss-120b fits on one 80 GB GPU such as an NVIDIA H100, and gpt-oss-20b runs with 16 GB of memory. Ollama's downloads are 65 GB and 14 GB.

Has OpenAI released a newer open-weight model? Not a general one as of Nov 6, 2026. After gpt-oss, OpenAI's open releases on Hugging Face were the gpt-oss-safeguard classifiers (Oct 29, 2025) and smaller research releases.

Can gpt-oss see images? No. Both gpt-oss models take and produce text only.