OpenAI ships its first open models since GPT-2

In August 2025, OpenAI released gpt-oss: two open-weight language models you can download, run, and modify yourself. It's the company's first open release since GPT-2 back in 2019, and it lands under an Apache 2.0 license that permits commercial use.

The models come in two sizes. The smaller one can run on a machine with about 16GB of RAM; the larger needs a high-end GPU or dedicated hardware. On several benchmarks they score in the neighborhood of OpenAI's smaller hosted models, which is respectable for something that runs locally with no per-token API bill.

For a small or mid-size business, the takeaway is optional leverage, not a mandate. You don't need to run your own model. But if you handle sensitive data you'd rather not send to a third party, or you want predictable costs on a repetitive task like summarizing or classifying documents, a local open model is now a real option worth pricing out against the usual API subscription.