OpenAI, the developer of ChatGPT, introduced a partially open-source artificial intelligence (AI) model on the 5th (local time).
On this day, OpenAI announced that it had released two AI models, 'GPT-oss-120b' and 'GPT-oss-20b', on the AI software sharing platform Hugging Face.
Both models are 'open-weight' models similar to Meta's Llama. 'Open weight' is characterized by not being fully open-source, but allows users to download the model's weights (the key values learned by the model) to run on personal computers or clouds and customize them for their use.
'Weight' refers to the numerous parameters within an AI model, and an AI defined as an 'open weight' model has numerous weight values that were acquired and adjusted during learning made public. However, since it is not completely open-source, the dataset used for training is not disclosed.
This is the first time OpenAI has launched an open-weight model since GPT-2 in 2019. GPT-2 was the software that served as the foundation for the early ChatGPT.
Most of OpenAI's models are closed, making user modifications impossible, with technical foundational information also disclosed only to a limited extent. OpenAI has maintained a strategy of recovering enormous development expenses through the protection of training datasets and the monetization of high-performance models.
Both models released this time are characterized by being relatively small and efficient. The high-performance model, 120b, can operate with one 80-gigabyte (GB) graphics processing unit (GPU), while the 20b can be run on a regular notebook with 16GB of memory, according to OpenAI. The two models can generate text based on user prompts and can perform complex tasks such as code writing or online information searches. However, they do not have image or video generation capabilities.
The two models released this time come six months after China's DeepSeek launched an open-source model 'RI' that competes with some of OpenAI's models in performance in January of this year. When the DeepSeek phenomenon swept through the tech industry early this year, OpenAI CEO Sam Altman noted that OpenAI was discussing the release of open-weight models.
OpenAI said, "We hope that the models released this time will be used by individuals, corporations, and government agencies to directly tune and run AI systems tailored to them."