Vibe Check: OpenAI Drops Two New Open-weight Models

Was this newsletter forwarded to you? Sign up to get it in your inbox.

If your life’s dream is to be able to vibe code on a deserted island, or in a cabin in the woods during a power outage, or on a long-haul flight with no WiFi, rejoice! Dreams do come true.

Today, OpenAI is releasing its first open-weight models since GPT-2 in 2019: gpt-oss-120b and gpt-oss-20b. Because they’re open-weight, you can run them yourself, on your laptop or on your own cloud, instead of in OpenAI’s cloud through its API. So you can take them with you anywhere—to a cabin in the woods—and use them at any time—at 40,000 feet with no WiFi.

The 20b is the small version that’s tiny enough for you to run on a MacBook. It’s about as powerful for coding and other reasoning tasks as o3-mini. Let me repeat: You can now run an o3-mini-level model on a MacBook.

The 120b is the more powerful model designed to run on a single GPU in your private cloud. According to OpenAI’s internal benchmarks, it matches or exceeds o4-mini on reasoning tasks like competition coding, general problem solving, and tool calling.

Usually I like to be hands-on with vibe checks, but unfortunately I couldn’t try these models before their release. However, based on OpenAI’s internal benchmarks, both of them appear to push the current frontier for other open models of their size, like Meta’s Llama and Alibaba’s Qwen. As Greg Brockman, OpenAI cofounder and president, said during a call with journalists yesterday, “The team cooked with this one.”

Both models are being released today for free under an Apache 2.0 open-source software license, so you can use, modify, or sell them yourself.

The big questions are: What are these models useful for? And why? What’s the strategy?

What are these models useful for?