gpt-oss
OpenAI's Apache 2.0 reasoning models from 2025: 117B and 21B mixtures of experts that fit one 80 GB GPU or a 16 GB machine.
In short. gpt-oss is OpenAI's pair of downloadable AI models: a larger one that fits on a single 80 GB data-center GPU and a smaller one that runs in about 16 GB of memory. They think step by step, can call tools, and use the Apache 2.0 license. They are older than most models on this site, but they remain widely used.
Summary
OpenAI released gpt-oss on 5 August 2025 as two mixture-of-experts reasoning models under the Apache 2.0 license. gpt-oss-120b has 117 billion parameters with 5.1 billion active per token; gpt-oss-20b has 21 billion with 3.6 billion active. The larger model has 128 experts per layer with 4 active per token. Both have a context length of 131,072 tokens and read and write text only.
OpenAI post-trained the models with MXFP4 quantization of the expert weights, which is why the checkpoints are small: about 65 GB for 120b, which runs on a single 80 GB GPU, and about 14 GB for 20b, which OpenAI says runs within 16 GB of memory. The models offer configurable reasoning effort (low, medium, high), expose their full chain of thought, and support function calling, web browsing, Python code execution and structured outputs. They must be used with OpenAI's "harmony" response format. The license is Apache 2.0, accompanied by a short usage policy that asks you to comply with applicable law.
They are older than most entries in this catalog, but gpt-oss is still OpenAI's main open-weight language-model family. Its later open releases on Hugging Face are a safety-classification variant (gpt-oss-safeguard), a research model and a personal-data filter, not general-purpose successors. Hugging Face showed about 5.9 million downloads of gpt-oss-20b in the 30 days to 8 November 2026, making it one of the most downloaded open models.
Good for
- A strong reasoning model on one 80 GB GPU (120b) or a 16 GB laptop or GPU (20b)
- Local agents and tool calling, with function calling, browsing and Python built into the training
- Commercial products, since Apache 2.0 imposes no revenue or hosting conditions
- A well-supported default, since nearly every inference tool already handles gpt-oss
Watch out for
- Text only, with no image, audio or video input
- It is more than a year old, and newer models in this catalog lead on many benchmarks
- It must be used with the harmony response format; Transformers' chat template applies it, but hand-written prompts will not work correctly
- The chain of thought is raw and is not meant to be shown to end users
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Estimates for the gpt-oss-20b variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.
| Variants | Parameters | BF16 | FP8 / INT8 | 4-bit (Q4) | Fits on |
|---|---|---|---|---|---|
| gpt-oss-20b MXFP4 | 21B 3.6B active | 48 GB | 24 GB | 15 GB | Laptop or phone |
| gpt-oss-120b MXFP4 | 117B 5.1B active | 269 GB | 135 GB | 72 GB | Workstation |
Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.
The license in plain words
One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.
See also
- Best open-source LLMs in 2026
- Best open-source models for coding and AI agents
- Best small open models you can run on a laptop
- Clef: Cloudflare's open decision models for agent workflows
- Giant open models now ship in 4-bit: what it means for memory planning
Frequently asked questions
- What is gpt-oss?
- gpt-oss is OpenAI's pair of downloadable AI models: a larger one that fits on a single 80 GB data-center GPU and a smaller one that runs in about 16 GB of memory. They think step by step, can call tools, and use the Apache 2.0 license. They are older than most models on this site, but they remain widely used.
- Can I use gpt-oss commercially?
- Yes. gpt-oss is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
- Is gpt-oss open source?
- gpt-oss is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
- Can I run gpt-oss on my own computer?
- Yes. The smallest variant, gpt-oss-20b, needs about 15 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
- When was gpt-oss released?
- gpt-oss was released by OpenAI on August 5, 2025.
- How long a context window does gpt-oss have?
- gpt-oss supports a context window of up to 128K tokens.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- gpt-oss-120b model card (Hugging Face) huggingface.co
- gpt-oss-20b model card (Hugging Face) huggingface.co
- gpt-oss repository (GitHub) github.com
Release date and facts last verified November 8, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| Qwen3.8-27B Alibaba Qwen | 27.8B | 256K | Apache 2.0 (Permissive) | 19 GB Gaming GPU | |
| Gemma 4 Google DeepMind | 5.1B – 30.7B | 256K | Apache 2.0 (Permissive) | 3.52 GB Laptop or phone | |
| Nemotron 3.5 Lightning NVIDIA | 31.6B 3B active | 1M | OpenMDW-1.1 (Permissive) | 22 GB Gaming GPU | |
| Clef Cloudflare | 9.41B – 27.4B | 64K | Apache 2.0 (Permissive) | 6.49 GB Laptop or phone |