Gpt-oss-20b
OpenAI's smaller open-weight reasoning model for local and low-latency use
gpt-oss-20b is the smaller of two open-weight reasoning models released by OpenAI in August 2025. It has 21 billion parameters with 3.6 billion active per token, runs within 16GB of memory, and is aimed at low-latency, local and specialised tasks including agentic tool use.
- 21B total parameters with 3.6B active per token
- Runs within 16GB of memory after MXFP4 quantization
- Weights released under a permissive Apache 2.0 licence
- Reasoning effort adjustable across low, medium and high
- Fine-tunable on consumer hardware
What it is
gpt-oss-20b is an open-weight language model built by OpenAI for reasoning and agentic work. It is the smaller member of the gpt-oss pair, sized at 21 billion parameters of which 3.6 billion are active for each token generated. OpenAI positions it for lower latency and for running locally or on specialised tasks, while the larger gpt-oss-120b targets production use. The model handles function calling, web browsing, Python execution and structured outputs, and its reasoning effort can be set to low, medium or high. Weights, inference code, tool environments and tokenizers were published under Apache 2.0.
How it came about
OpenAI announced gpt-oss-20b alongside gpt-oss-120b as its first open-weight models in some time, describing them as designed for reasoning, agentic tasks and developer use. The model card was posted to arXiv on 8 August 2025 with more than 120 listed authors, and the weights were published on the Hugging Face Hub. Both models were trained in the harmony response format and are meant to be used only with that format.
Inside the model
The model uses a mixture-of-experts transformer design, where only a subset of the network is active for each token. Training combined large-scale distillation with reinforcement learning. Post-training applied MXFP4 quantization to the mixture-of-experts weights, which is what lets the 21B model fit within 16GB of memory; all reported evaluations used that same quantization. It was trained in the harmony response format, which renders roles and instructions explicitly. OpenAI says the model can be fine-tuned on consumer hardware.
What it can do
OpenAI reports strong results across mathematics, coding and safety benchmarks, though specific scores are not given here. The model exposes its full chain of thought, which OpenAI says helps debugging and trust but should not be shown to end users. Reasoning effort is adjustable across three levels: low for fast dialogue, medium for balance, high for deep analysis. Native agentic skills cover browsing, function calling with schemas, Python execution and structured outputs.
How to use it
Weights are downloadable from the Hugging Face Hub under a permissive Apache 2.0 licence, with no copyleft or patent restrictions. It runs through Transformers, vLLM, Ollama and LM Studio, and can be served as an OpenAI-compatible webserver. The Transformers chat template applies the required harmony format automatically.
Safety and testing
The model card states that the models achieve strong results on safety benchmarks alongside maths and coding. No further safety evaluation details, red-teaming results or incidents are described here.
Why it matters
OpenAI released the weights, inference implementations, tool environments and tokenizers to encourage broad use and further research. The pair is presented as pushing accuracy against inference cost, with the smaller model aimed at consumer hardware and fine-tuning outside large data centres.
Questions people ask about Gpt-oss-20b
What is gpt-oss-20b?
It is the smaller of OpenAI's two open-weight reasoning models, released in August 2025. It has 21 billion parameters with 3.6 billion active per token and is aimed at low-latency, local and specialised tasks, including agentic work such as browsing and function calling.
How is it different from gpt-oss-120b?
gpt-oss-120b is the larger model, with 117 billion parameters and 5.1 billion active, intended for production and high reasoning work on a single 80GB GPU. gpt-oss-20b is smaller, runs within 16GB of memory, and targets lower latency and local use.
Can I use it commercially?
Yes. OpenAI released the weights under an Apache 2.0 licence, which it describes as permissive, with no copyleft restrictions or patent risk, suitable for experimentation, customisation and commercial deployment.
What hardware do I need?
After MXFP4 quantization of the mixture-of-experts weights, the model fits within 16GB of memory and can run on consumer hardware through tools such as Ollama or LM Studio. OpenAI says it can also be fine-tuned on consumer hardware.
Why does it need the harmony format?
Both gpt-oss models were trained on OpenAI's harmony response format and must be used with it, as they will not work correctly otherwise. The Transformers chat template applies the format automatically.
Good reading and sources
There is no English Wikipedia article on Gpt-oss-20b yet, so the facts on this page come from the sources listed above: huggingface.co/openai/gpt-oss-20b, arxiv.org/abs/2508.10925. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.