gpt-oss-20b: OpenAI's Open Weights, and the Format You Must Use

What gpt-oss-20b is

gpt-oss-20b is the smaller of OpenAI's two open-weight releases — 21 billion parameters with 3.6 billion active, published under Apache 2.0 with a 128K token context window. OpenAI position it for lower-latency, local and specialised use, with the larger gpt-oss-120b (117 billion parameters, 5.1 billion active) aimed at production work and sized deliberately to fit a single 80GB GPU such as an H100 or MI300X. Reasoning effort is configurable on both, and both are designed for reasoning, agentic tasks and developer use.

The harmony response format

Both gpt-oss models were trained on OpenAI's harmony response format, and OpenAI state that they should only be used with it because they will not work correctly otherwise. OpenAI publish the harmony specification as a separate repository.

LMArena and price

On LMArena's text leaderboard gpt-oss-20b has 1317 Elo (rank 270); by category, 1368 Elo in Coding (rank 252), 1335 Elo in Math (rank 229), 1315 Elo in Expert (rank 252) and 1240 Elo in Creative Writing (rank 307). It is roughly $0.03 per million input tokens and $0.13 output through third-party providers.

Sizing

gpt-oss-120b, the larger sibling, has 117 billion parameters with 5.1 billion active and is sized to fit a single 80GB GPU. Both are Apache 2.0.

Features listed on the card

OpenAI list configurable reasoning effort at low, medium and high, set in the system prompt with a line such as "Reasoning: high". The model exposes its full chain of thought for debugging, which OpenAI state is not intended to be shown to end users. It supports function calling, web browsing, Python code execution and Structured Outputs, and can be fine-tuned; OpenAI state gpt-oss-20b can be fine-tuned on consumer hardware, while gpt-oss-120b needs a single H100 node. The mixture-of-experts weights were post-trained with MXFP4 quantization, which lets gpt-oss-20b run within 16GB of memory and gpt-oss-120b on one 80GB GPU; all of OpenAI's evaluations used the same quantization.

Running it

The model runs with Hugging Face Transformers, where the chat template applies the harmony format automatically; calling model.generate directly requires applying harmony manually or through the openai-harmony package. OpenAI document serving through vLLM, Transformers Serve and Ollama as OpenAI-compatible endpoints, and downloads through LM Studio and the Hugging Face CLI. Reference implementations for PyTorch and Triton are in the gpt-oss repository, and the model card is published on arXiv as 2508.10925. The weights are released under Apache 2.0, which OpenAI describe as carrying no copyleft restrictions or patent risk.

Where to get it

OpenAI publish both models in a gpt-oss collection on the Hugging Face Hub; the original checkpoint files download with the Hugging Face CLI by including the original folder. The models can be pulled in Ollama as gpt-oss:20b and in LM Studio as openai/gpt-oss-20b, and tried in a browser at gpt-oss.com. OpenAI maintain a list of inference partners and resources in the gpt-oss GitHub repository, alongside guides in the OpenAI Cookbook for running the models with Transformers, vLLM and Ollama. The harmony specification is a separate repository, github.com/openai/harmony.

Official Resources