OpenAI has unveiled its inaugural open-weight models in more than five years, introducing the language models gpt-oss-120b and gpt-oss-20b. These models can be executed locally on consumer devices and fine-tuned for particular tasks, representing a shift from OpenAI’s recent emphasis on proprietary releases. CEO Sam Altman conveyed enthusiasm about making these models, the culmination of extensive research, accessible to the public to enhance AI accessibility. Both models are available for free download on Hugging Face. The last open-weight model offered by OpenAI was GPT-2, launched in 2019.
Open-weight models are characterized by having their “weights” publicly available, enabling anyone to scrutinize the internal parameters and comprehend information processing. OpenAI co-founder Greg Brockman characterized the release as supplementary to the company’s paid services, such as its application programming interface. In contrast to ChatGPT, gpt-oss models are capable of functioning offline and within firewalls.
The gpt-oss models utilize chain-of-thought reasoning, a technique first implemented in OpenAI’s o1 model, which incorporates multiple steps to generate responses. While these models are text-only and not multimodal, they can browse the web, leverage cloud-based models, run code, and act as AI agents. The smaller model, gpt-oss-20b, is sufficiently compact to operate on devices with over 16 GB of memory.
These models are made available under the Apache 2.0 license, permitting commercial use, redistribution, and incorporation into other licensed software. Comparable releases from Alibaba’s Qwen and Mistral also utilize Apache 2.0.
Initially announced in March, the launch was postponed for further safety assessments. Open-weight models carry more risk than their closed counterparts due to fewer usage limitations, which could lead to potential misuse. OpenAI performed tailored evaluations on the open-weight option to measure misuse risks. Safety researcher Eric Wallace mentioned that internal fine-tuning on risk areas indicated the model did not reach high risk levels according to OpenAI’s preparedness framework.
