gpt-oss

0:25 gpt-oss is OpenAI's family of open-weight models, released on Aug 5, 2025: gpt-oss-120b and gpt-oss-20b, reasoning models that anyone can download, run and change for free under the Apache 2.0 license. The 20b runs within 16 GB of memory, so it fits a good laptop or desktop; the 120b fits a single 80 GB GPU such as an Nvidia H100. Both have 128K tokens of context (Ollama), and OpenAI has released no newer general gpt-oss model as of Nov 6, 2026.

0:00 Pick gpt-oss for a capable reasoning model you can run privately on your own hardware, with a simple permissive license and tool use (browsing, Python, function calls) built in. It reads only text, though: for image input, or for the newest open models, look at Gemma, Qwen, DeepSeek, Mistral or Llama. For OpenAI's best models, use ChatGPT or the API instead. See Run AI Locally.

Learn more: gpt-oss on chatgpt.vi.ki

Quick facts

Models and hardware

Sources: gpt-oss-120b and gpt-oss-20b model cards and Ollama's gpt-oss page, read Nov 6, 2026.

0:20 Both are mixture-of-experts models, so only a small part of each runs for every word, which keeps them fast. OpenAI trained them with their expert weights already in a 4-bit format (MXFP4), which is how the 120b fits on one GPU and the 20b in 16 GB. A Mac with 16 GB or more of unified memory, or a PC with a 16 GB graphics card, can run the 20b.

Features

1:00 Reasoning effort. Both models think before answering, and you choose low, medium or high effort in the system prompt ("Reasoning: high"), trading speed for accuracy, per the model cards.

Tools and agents. OpenAI built them for agent work: web browsing, running Python and calling functions you define, per OpenAI's model card. The gpt-oss repository includes reference browser, Python and file-patching tools.

Harmony format. The models expect OpenAI's "harmony" chat format and do not work correctly without it; Transformers, Ollama and LM Studio apply it for you (GitHub).

Fine-tuning. The Apache 2.0 license allows changing and selling the models, and OpenAI says the 20b can be fine-tuned on consumer hardware.

gpt-oss-safeguard. A later pair of models built on gpt-oss that classify text against a safety policy you write, for moderation and content labeling, also under Apache 2.0.

Where to run gpt-oss

• On your computer: ollama run gpt-oss:20b in Ollama, or lms get openai/gpt-oss-20b in LM Studio (GitHub).
• On a server: vLLM gives an OpenAI-compatible endpoint; Transformers works too.
• Hosted: many providers serve it by the token; OpenRouter lists several, and Hugging Face Inference Providers route to them. Ollama also offers cloud tags (gpt-oss:120b-cloud) that run on its servers.

gpt-oss vs the alternatives

• Qwen: Alibaba's open models come in many more sizes and read images; a common choice for local coding.
• Gemma: Google's open models cover phones to workstations and take image input.
• DeepSeek: much larger MIT-licensed models with 1 million tokens of context, best on a server or through its cheap API.
• Mistral: Ministral 3 and Mistral Small 4 under Apache 2.0, from a European maker.
• Llama: Meta's open models, with a more restrictive license and no release since Apr 2025.

Frequently asked questions

Is gpt-oss free? Yes. OpenAI's gpt-oss-120b and gpt-oss-20b are free to download, use, change and sell under the Apache 2.0 license; you pay only for your own hardware or a hosting provider.

Can I run gpt-oss on my laptop? gpt-oss-20b runs within 16 GB of memory, so a laptop or desktop with 16 GB of unified memory or a 16 GB graphics card can run it; gpt-oss-120b needs an 80 GB GPU.

Is gpt-oss the same as ChatGPT? No. gpt-oss is a pair of open-weight models from Aug 2025 that you run yourself; ChatGPT runs OpenAI's newer, larger closed models.

Does gpt-oss support images? No. gpt-oss-120b and gpt-oss-20b take text only; open models such as Gemma, Qwen and Llama 4 accept images.

What license is gpt-oss under? gpt-oss uses the Apache 2.0 license, which allows commercial use and changes with no user-count limit.

Sources

• gpt-oss-120b, Hugging Face, read Nov 6, 2026
• gpt-oss-20b, Hugging Face, read Nov 6, 2026
• gpt-oss-safeguard-20b, Hugging Face, read Nov 6, 2026
• gpt-oss-120b & gpt-oss-20b Model Card, OpenAI, Aug 8, 2025
• openai/gpt-oss, GitHub, read Nov 6, 2026
• gpt-oss, Ollama, read Nov 6, 2026
• gpt-oss-120b, Artificial Analysis, read Nov 6, 2026
• gpt-oss: Everything You Need to Know in Under 2 Minutes, BlueSpork, Aug 8, 2025 (video)
• ChatGPT-OSS Is Here And Can Run Locally On Your PC, CNET, Aug 11, 2025 (video)