OpenAI has made a significant announcement by releasing its first open-weight models since 2019, namely the gpt-oss-120b and gpt-oss-20b. These models can operate on consumer devices and can be customized for specific applications, indicating a strategic shift from OpenAI’s recent focus on proprietary software to a more open model landscape.
Sam Altman, OpenAI’s CEO, expressed excitement about making this technology accessible to a broader audience, which he believes can lead to greater benefits. The models are available for free download on Hugging Face, a well-known platform for hosting AI tools. The last time OpenAI provided an open-weight model to the public was with GPT-2.
Open-weight models allow users to view the underlying parameters that dictate how the model processes information. This level of transparency does not undermine OpenAI’s paid models; instead, co-founder Greg Brockman suggests they complement existing services, as these models can function without internet access.
The gpt-oss models use chain-of-thought reasoning, a technique OpenAI first introduced with its o1 model. Although these new models are text-only, they feature capabilities such as web browsing, execution of code, and navigation of software as an AI agent. The smaller gpt-oss-20b can run effectively on devices with more than 16 GB of memory.
These models are distributed under the Apache 2.0 license, which allows for commercial use and redistribution. OpenAI ensured a thorough safety evaluation of these models, given the potential risks associated with open-weight releases. They tested how the model could be misused if it fell into the wrong hands and found that risks were manageable.
Performance comparisons show that the gpt-oss models have strong benchmark scores, with the gpt-oss-120b performing closely to OpenAI’s proprietary o3 and o4 Mini models, and even exceeding them in some areas. Furthermore, the release occurs amid a competitive landscape in the AI sector, particularly with companies like Meta, which has also pursued advancements in open-source AI.
Despite the potential risks, Altman emphasized OpenAI’s mission to foster innovation around open-weight models within the United States, aiming to create an AI ecosystem grounded in democratic values and accessible to everyone.
