gpt-oss (gpt-oss-120b and gpt-oss-20b)

What it is

gpt-oss is a set of open-weight reasoning language models released by OpenAI on 5 August 2025. They come in two sizes, gpt-oss-120b and gpt-oss-20b.^[1] "Open-weight" means the trained parameters are published so that others can download, run and adapt the model. It does not necessarily mean that the training data or full training code is public.

Origins

The release came two days before GPT-5, and OpenAI compared the models' performance with its own o4-mini and o3-mini models.^[1] Its OpenAI-hosted documentation lists a 131,072-token context window and a knowledge cutoff of 1 June 2024.^[3]

Main features

gpt-oss-120b has 117 billion total parameters, with 5.1 billion active for each token. gpt-oss-20b has 21 billion total parameters and 3.6 billion active per token. Both use a mixture-of-experts (MoE) design, in which only part of the network is activated for each token.^[2]

The mixture-of-experts weights are quantised to MXFP4, a low-precision number format that shrinks their memory footprint. As a result gpt-oss-120b fits on a single 80 GB GPU such as an NVIDIA H100 or AMD MI300X, and gpt-oss-20b runs within 16 GB of memory.^[2]

Both models are released under the Apache 2.0 license, which allows commercial use.^[1][2] In general terms, the license lets people use, modify and redistribute the software and the models, subject to the conditions of the license text. Reasoning effort is configurable as low, medium or high. Developers can access the full chain of thought, although it is not intended for end users.^[2]

The models were trained on OpenAI's harmony response format and support function calling, web browsing, Python execution and Structured Outputs.^[2] Weights are distributed on Hugging Face, with support in Transformers, vLLM, Ollama and LM Studio.^[2]

A reasoning model spends part of its output on intermediate steps before giving a final answer. Function calling lets the model ask the developer software to run a specified function, such as a lookup, and then use the result in its reply.^[2]

Place in the development

OpenAI's open-weight line later added gpt-oss-safeguard-20b, a safety-classification variant, on 29 October 2025.^[4]

The models show OpenAI releasing weights that others can run themselves, alongside its hosted models. For readers, the key contrast is that a hosted model runs on OpenAI's servers, while an open-weight model can be downloaded and run on hardware the user controls.