Openai has just abandoned its first open model in five years. Two language models of GPT-OSS-1220B and GPT-OSS-20B can run locally on consumer devices and are used for specific purposes. For Openai, their shift away from its latest strategy focused on proprietary versions as the company moves towards a broader, more open AI model toward proprietary versions.
“We are excited to be able to make this model, the result of a billion-dollar research that can put AI in the hands of most people,” Openai CEO Sam Altman said in an emailed statement. Both GPT-OSS-1220B and GPT-OSS-20B are available for free on Hugging Face, a popular AI tool hosting platform. The last open weight model released by OpenAI was GPT-2, back in 2019.
The fact that the open weight model is unique is that its “weight” is publicly available, meaning anyone can peek inside parameters to understand how it processes information. Instead of undermining OpenAI’s proprietary model with free options, co-founder Greg Brockman sees this version as a “supplement” to the company’s paid services, just like the application programming interface that many developers currently use. “Open models have a very different set of advantages,” Brockman said in a briefing with reporters. Unlike Chatgpt, you can run GPT-oss models without an Internet connection and behind a firewall.
Both GPT-ss models use thoughtful reasoning methods, and OpenAI was first deployed in its O1 model last fall. This method not only needs to output output, but also has generated AI tools that can answer prompts through multiple steps. These new text-only models are not multimodal, but they can browse the web, call cloud-based models to help with tasks, execute code, and software that serves as AI agents. The smaller GPT-OSS-20B of these two models is compact enough to run locally on consumer devices with over 16 GB of memory.
Two new models from OpenAI are available under Apache 2.0 license, a popular choice for open weight models. With Apache 2.0, the model can be used for commercial purposes, redistributed, and used as part of other licensed software. Open weight models issued from Alibaba’s Qwen and Mistral are also running under Apache 2.0.
Publicly announced in March that the release of these open models was initially postponed for further security testing. Releasing an open weight model can be more dangerous than a closed version, as it removes the barriers to using the tool, and anyone can try to use the GPT-oss version for unexpected purposes.
In addition to evaluating OpenAI usually runs on its proprietary models, the startup has customized open weight options to see the “bad actors” who downloaded the tool may abuse it. “In fact, we fine-tuned these risk areas internally and measured the height we can push them,” said Eric Wallace, security researcher at OpenAI. In OpenAI’s test, by its preparation framework, the open model did not reach high risk.
