Gpt-oss-120b
OpenAI's open-weight reasoning model that fits on one 80GB GPU
gpt-oss-120b is an open-weight reasoning model from OpenAI, released in August 2025 alongside a smaller sibling, gpt-oss-20b. It uses a mixture-of-experts transformer and is aimed at production, general-purpose and agentic work, with weights published under Apache 2.0.
- 117B total parameters with 5.1B active per token
- Runs on a single 80GB GPU such as an NVIDIA H100 or AMD MI300X
- Weights released under Apache 2.0
- Reasoning effort adjustable across low, medium and high
- MXFP4 quantization applied to mixture-of-experts weights in post-training
What it is
gpt-oss-120b is the larger of two open-weight reasoning models released by OpenAI. It has 117 billion parameters in total, of which 5.1 billion are active for each token generated, and is built for production, general-purpose and high-reasoning tasks. OpenAI says it targets work that fits within a single 80GB GPU such as an NVIDIA H100 or AMD MI300X. The model handles agentic work such as function calling, web browsing and Python code execution, and its reasoning effort can be set to low, medium or high.
How it came about
OpenAI published the model card for gpt-oss-120b and gpt-oss-20b on arXiv on 8 August 2025, with a long author list led by Sandhini Agarwal and including Sam Altman and many other OpenAI staff. The paper describes both models as open-weight reasoning systems trained with large-scale distillation and reinforcement learning. Weights, inference implementations, tool environments and tokenizers were released together under Apache 2.0.
Inside the model
The model uses a mixture-of-experts transformer design, where only a subset of the network is active for each token. OpenAI says training combined large-scale distillation with reinforcement learning. During post-training, the mixture-of-experts weights were quantized to MXFP4, a compact number format, which is what allows the 117B model to run on one 80GB GPU. All reported evaluations used that same quantization. The model was trained on the harmony response format and must be run with it, since it does not work correctly otherwise. Fine-tuning is supported, and OpenAI says the large model can be fine-tuned on a single H100 node.
What it can do
OpenAI reports strong results across mathematics, coding and safety benchmarks, though no individual scores are given here. Reasoning effort can be tuned across three levels: low for fast general dialogue, medium for a balance of speed and detail, and high for deep analysis. The full chain of thought is exposed to developers for debugging, though OpenAI says it is not meant to be shown to end users. Native agentic features include function calling with defined schemas, built-in web browsing, Python execution and structured outputs.
How to use it
Weights are downloadable from the Hugging Face Hub under a permissive Apache 2.0 licence, which OpenAI describes as free of copyleft restrictions and patent risk. The model runs through Transformers, vLLM, PyTorch and Triton reference code, Ollama and LM Studio, and can be served as an OpenAI-compatible web server. The harmony response format is required for correct behaviour.
Safety and testing
The model card states that the models achieve strong results on safety benchmarks alongside mathematics and coding. Full chain-of-thought access is provided to help developers debug outputs and build trust. No further safety evaluation details are given here.
Why it matters
The release gave developers and researchers open weights for a large reasoning model under a permissive licence, along with inference code, tool environments and tokenizers. OpenAI frames the pair as enabling broad use and further research, and points to a community list of gpt-oss resources and inference partners.
Questions people ask about Gpt-oss-120b
What is gpt-oss-120b?
It is the larger of OpenAI's two open-weight reasoning models, released in August 2025. It has 117 billion parameters with 5.1 billion active at a time and is aimed at production, general-purpose and agentic tasks, with weights published under Apache 2.0.
What hardware do I need to run it?
OpenAI says the model fits on a single 80GB GPU such as an NVIDIA H100 or AMD MI300X, thanks to MXFP4 quantization of the mixture-of-experts weights applied during post-training. The smaller gpt-oss-20b runs within 16GB of memory.
Can I fine-tune it?
Yes. OpenAI says both models can be fine-tuned for specialised use cases, and that gpt-oss-120b can be fine-tuned on a single H100 node. The smaller gpt-oss-20b can even be fine-tuned on consumer hardware.
Why does it need the harmony format?
Both models were trained on OpenAI's harmony response format and must be used with it. OpenAI states the model 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-120b yet, so the facts on this page come from the sources listed above: huggingface.co/openai/gpt-oss-120b, arxiv.org/abs/2508.10925, openai.com/index/introducing-gpt-oss/. 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.