OpenAI Releases 'GPT-oss' Open Models
First Open Models Since GPT-2
OpenAI, the developer behind ChatGPT, has announced that it will release two free artificial intelligence (AI) models capable of mimicking human reasoning processes.
On August 5 (local time), OpenAI stated that it plans to make two models, 'GPT-oss-120b' and 'GPT-oss-20b', available on the AI software hosting platform Hugging Face.
These two models use an 'open-weight' system similar to Meta's Llama. 'Open-weight' is not fully open source, but it does disclose the parameters learned by the AI during training, allowing developers to customize the models. OpenAI expressed its hope that individuals, businesses, and government agencies will be able to directly adjust and operate AI systems tailored to their needs using these models.
This is the first time OpenAI has released open models since GPT-2 in 2019. GPT-2 was the foundational software for the early versions of ChatGPT. Most of OpenAI's models have been closed systems, which cannot be modified by users and have only limited technical information disclosed. Like other competitors, OpenAI has recouped significant development costs by protecting training data and monetizing high-performance models.
The two newly released models are considered relatively small and efficient. 'GPT-oss-120b' can run on a single 80-gigabyte (GB) graphics processing unit (GPU), and 'GPT-oss-20b' can operate even on a standard laptop equipped with 16GB of memory.
OpenAI's release of open-weight models appears to have been influenced by the Chinese AI startup DeepSeek. Earlier this year, DeepSeek launched its open AI models, including 'R1', which attracted significant attention from the global tech industry. In response, OpenAI CEO Sam Altman previously stated that OpenAI was also discussing open-weight releases and emphasized the need to establish a new open-source strategy.
© The Asia Business Daily(www.asiae.co.kr). All rights reserved.
