OpenAI takes on Meta with two landmark open-source models
Apache 2.0 licensed ftw.
OpenAI has open-sourced two new models under a highly permissive Apache 2.0 licence, in its first fully open-weight release since 2019.
gpt-oss-120b and gpt-oss-20b are both general purpose, text-only models that use a mixture-of-experts (MOE) approach, OpenAI said today.
- The 34-page model card is here
- The GitHub repository is here.
- Hugging Face
- Fireworks.ai
gpt-oss-120b was designed for production-grade, general-purpose uses and fits on a single H100 GPU; gpt-oss-20b for “low-latency and resource-constrained environments”; think faster inference at the edge.
The two new OpenAI open-source models were trained on the harmony response format for defining conversation structures, generating reasoning output and structuring function calls,” OpenAI said.
“If you are not using gpt-oss directly but through an API or a provider like HuggingFace, Ollama, or vLLM, you will not have to be concerned about this as your inference solution will handle the formatting,” it added.
Those building their own inference solutions will need to get familiar with the Harmony format (also Apache 2.0) or the new open-source OpenAI models will “not work correctly” their documentation emphasised.
They were trained to perform strongly on STEM, coding, and general knowledge, said OpenAI. CEO Sam Altman added: “We are excited for the world to be building on an open AI stack created in the United States, based on democratic values, available for free to all and for wide benefit.”
The smaller model can be run on a smartphone, added Altman.
“We're excited to make this model, the result of billions of dollars of research, available to the world to get AI into the hands of the most people possible. We believe far more good than bad will come from it.”
Lin Qiao, founder of Fireworks.ai (a platform for optimising model inference) was among those welcoming their release. She said of the new open-source OpenAI models that “what really makes them shine is the quality of training data and reinforcement learning tuning…”
“They support long context, multi-step tool use, and offer adjustable reasoning levels (low/mid/high). Think o3/o4-mini performance, but open,” Qiao added: “You can use Fireworks supervised fine-tuning on these models today and reinforcement fine-tuning is coming shortly!”
Their release comes as The Information reported that OpenAI is building a full-blown consulting arm staffed with its own engineers...
Views on the release? Get in touch