OpenAI releases lower-cost models to rival Meta, Mistral and DeepSeek

CNBC

9 hours ago

OpenAI on Tuesday released two open-weight language models for the first time since it rolled out GPT-2 in 2019.
The text-only models are called gpt-oss-120b and gpt-oss-20b, and are designed to serve as lower-cost options that developers, researchers and companies can easily run and customize, OpenAI said.
An artificial intelligence model is considered open weight if its parameters, or the elements that improve its outputs and predictions during training, are publicly available. Open-weight models can offer transparency and control, but they are different from open-source models, whose full source code becomes available for people to use and modify.
Several other tech companies, including Meta, Microsoft-backed Mistral AI and the Chinese startup DeepSeek, have also released open-weight models in recent years.
"It's been exciting to see an ecosystem develop, and we are excited to contribute to that and really push the frontier and then see what happens from there," OpenAI President Greg Brockman told reporters during a briefing.
The company collaborated with Nvidia, Advanced Micro Devices, Cerebras, and Groq to ensure the models will work well on a variety of chips.
"OpenAI showed the world what could be built on Nvidia AI — and now they're advancing innovation in open-source software," Nvidia CEO Jensen Huang said in a statement.
The release of OpenAI's open weight models has been highly anticipated, in part because the company repeatedly delayed the launch.
In a post on X in July, OpenAI CEO Sam Altman said the company needed more time to "run additional safety tests and review high-risk areas." That came after a separate post weeks earlier, where Altman said the models would not be released in June.
OpenAI said Tuesday that it carried out extensive safety training and testing on its open-weight models.
It filtered out harmful chemical, biological, radiological and nuclear data during pre-training, and it mimicked how bad actors could try to fine-tune the models for malicious purposes. Through this testing, OpenAI said it determined that maliciously fine-tuned models were not able to reach the "high capability" threshold in its Preparedness Framework, which is its method for measuring and protecting against harm.
The company also worked with three independent expert groups who provided feedback on its malicious fine-tuning evaluation, OpenAI said.
OpenAI said people can download the weights for gpt-oss-120b and gpt-oss-20b on platforms like Hugging Face and GitHub under an Apache 2.0 license. The models will be available to run on PCs through programs such as LM Studio and Ollama. Cloud providers Amazon, Baseten and Microsoft are also making the models available.
Both models can handle advanced reasoning, tool use and chain‑of‑thought processing, and are designed to run anywhere — from consumer hardware to the cloud to on-device applications.
Users can run gpt-oss-20b on a laptop, for instance, and use it as a personal assistant that can search through files and write, OpenAI said.
"We're excited to make this model, the result of billions of dollars of research, available to the world to get AI into the hands of the most people possible," Altman said in a statement Tuesday.