OpenAI releases low-cost model to compete with Meta (META.US) and DeepSeek.

OpenAI released its first batch of open-weight language models since the launch of GPT-2 in 2019 on Tuesday. These two pure text models are named gpt-oss-120b and gpt-oss-20b, aimed at providing developers, researchers, and businesses with easier to run and customizable low-cost options. When the parameters of an artificial intelligence model (the elements that optimize the output and predictive abilities during the training process) are publicly available, the model is considered open-weight. While these models provide transparency and control, they are different from open-source models, which allow users to fully access and modify the source code. In recent years, companies such as Meta, Microsoft-backed Mistral AI, and Chinese startup DeepSeek have also released open-weight models. OpenAI President Greg Brockman said in a briefing, "It's exciting to witness the growth of this ecosystem, and we are honored to be part of it and push the boundaries of technology to observe its development." The company has partnered with Nvidia, AMD, Cerebras, and Groq to ensure that the models can run stably on a variety of chips. "Nvidia has shown the possibilities of AI technology, and now they are driving innovation in the open-source software field," said Nvidia CEO Jensen Huang in a statement. The release was highly anticipated, in part because the company had delayed the launch multiple times before. OpenAI CEO Sam Altman wrote on the X Platform in July that more time was needed "for additional safety testing and reviewing high-risk areas," and he had previously stated in June that the model would not be released that month. OpenAI announced on Tuesday that the open-weight models had undergone comprehensive security training and testing. Harmful chemical, biological, radioactive, and nuclear weapons-related data were removed during the pre-training phase, and simulations of malicious actors' potential model fine-tuning were conducted. Tests showed that maliciously fine-tuned models could not reach the high capability threshold set by the "alignment framework," which is the company's assessment system for measuring and preventing harm. OpenAI also revealed that three independent expert teams had been invited to provide feedback on the assessment of malicious fine-tuning. Users can download model weights under the Apache 2.0 license on platforms like Hugging Face and GitHub and run them on personal computers using programs like LM Studio and Ollama. Amazon Cloud, Baseten, and Microsoft Cloud will also provide model services. Both models have advanced reasoning, tool calling, and thought chain processing capabilities, suitable for various environments from consumer-grade hardware to cloud and terminal devices. For example, gpt-oss-20b can run on laptops as a personal assistant that can retrieve files and assist in writing. "We are honored to open up this multi-billion-dollar research achievement to the world, allowing as many people as possible to master AI technology," Altman said in Tuesday's statement.