OpenAI’s new gpt-oss models come in two sizes: 120 billion and 20 billion parameters, or the statistical relationships learned by a model during training. Generally speaking, the higher the parameter count, the more capable the model.
“We believe this is the best and most usable open model in the world,” OpenAI CEO Sam Altman said in a post on X.
The last time OpenAI released an open model was in 2019, with GPT-2. But GPT-2 was fully open — making it truly open-source — albeit done in stages over months. OpenAI didn’t say whether gpt-oss will eventually be open-source as well.
Gpt-oss is a text-only, open-weights model, meaning the user can use and fine-tune the model but not know how it was trained or what data it was trained on.
Without knowing the data that went into the model, companies do not get full transparency, which could add risks to companies in financial services, healthcare and other highly regulated industries.
Advertisement: Scroll to Continue
For example, a healthcare company might want to fully vet a model’s training process before deploying it on personal patient data.
“If only open weights are available, developers … lack the ability to meaningfully evaluate biases, limitations, and societal impacts,” according to the Prompt Engineering and AI Institute.
However, OpenAI is granting access under the Apache 2.0 license, which gives the user “perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable copyright” to produce and distribute it.
“These open models also lower barriers for emerging markets, resource-constrained sectors, and smaller organizations that may lack the budget or flexibility to adopt proprietary models,” according to OpenAI.
Users can download GPT-OSS on Hugging Face or view on GitHub.
Read more: OpenAI Adds 200 Million Weekly Users in 4 Months
The Open AI Model Race
Major contenders for dominance in the open-source arena are Meta, with its Llama flagship of models but with usage and redistribution restrictions; France’s Mistral AI; and Chinese providers such as DeepSeek and Alibaba for most of its Qwen models.
Altman called gpt-oss “a big deal,” with “strong real-world performance” comparable to o4-mini. (The o, or omni, series comprises OpenAI’s reasoning models — a new family of models after its GPT series.)
For example, in the MMLU benchmark that tests how well LLMs perform a broad range of academic and professional tasks, gpt-oss is not far behind OpenAI o3 and o4-mini in performance.
OpenAI said the gpt-oss 120-billion parameter model can run on the user’s own computer and the smaller model can be run on a smartphone. Typically, AI models especially larger ones, run in the cloud.
While releasing an open model means bad actors can use it for evil, Altman said the company thinks “far more good than bad will come from it.” This is a departure from 2019, when OpenAI released GPT-2 in stages, fearful that it would used for ill.
The rest of OpenAI’s models are closed and proprietary. Its rivals have released varying degrees of open models: Google has open-weight, not open-source, models like Gemma. Anthropic doesn’t have an open model. Microsoft open-sourced its Phi models. Amazon’s models are proprietary.
Read more:
Apple AI Team Working on ChatGPT Rival
OpenAI Raises New Funding to Hit $300 Billion Valuation
Microsoft Seeks to Extend Access to OpenAI Technology
