Updated:AWS and Microsoft have announced that they will provide OpenAI's open weight language model gpt-oss series through their cloud service platforms. This open weight model can be used through Amazon Bedrock and Amazon SageMaker AI, or Azure AI Foundry, respectively.
In addition, Qualcomm also stated that the gpt-oss-16b model can be executed through a device equipped with the flagship Snapdragon computing platform and built-in 20GB of memory capacity.
OpenAI earlier announced a new open-weightLarge language models gpt-oss-120b and gpt-oss-20bThis is also the second time that OpenAI has launched a non-commercial open language model since the launch of the GPT-2019 language model in 2, symbolizing its response to the original intention of "letting artificial intelligence achievements benefit all mankind."
Unlike open-source models, GPT-OSS does not disclose its training data or code. Instead, it releases the weights generated after model training—the values assigned to the input data during the learning process. While the training method cannot be fully traced, users can still use it to infer, fine-tune, and even develop corresponding services.
According to Benjamin C. Lee, a professor of computer science at the University of Pennsylvania, open weight models are an option "between black-box business models and fully open source models." Developers can quickly import and apply powerful language models without high training costs.
Of the two models released this time, gpt-oss-120b boasts approximately 117 billion parameters, approaching the performance of OpenAI's commercial model O3, making it suitable for deployment on high-end hardware (a GPU with 80GB of RAM is recommended). In contrast, gpt-oss-20b requires only a computer with 16GB of RAM, making it more suitable for local computing on personal computers, such as offline programming assistance and text generation.
Both adopt the Apache 2.0 open source licensing terms, allowing developers to modify, deploy, and use them commercially, further lowering the threshold for AI model applications in enterprises or research.
In terms of functionality, the gpt-oss series supports advanced capabilities such as chain-of-thought reasoning and tool usage (such as executing Python code). Although it lacks multimodal input (such as images and voice), it can be supplemented by relevant functions through connecting to OpenAI's commercial services.
Notably, OpenAI conducted additional security testing and optimizations before releasing gpt-oss. They emphasized that these models will be tested as a pilot project, with subsequent feedback expected to determine whether to expand the open access experience. One of the initial partners is AI Sweden, Sweden's national artificial intelligence center, which will assist with future application deployment testing.
This move also coincided with Meta CEO Mark Zuckberg's revelation that "the release of open source models will be reduced in the future", thus forming a contrast with OpenAI.
