You can now download and run high-quality AI reasoning and safety models on infrastructure you control.
OpenAI published two open-weight reasoning models, gpt-oss-120b and gpt-oss-20b, and also released matching safety models called gpt-oss-safeguard-120b and gpt-oss-safeguard-20b. The weights are available under the Apache 2.0 license and are not served through the OpenAI API or ChatGPT.
The release is notable because OpenAI provided concrete sizing and purpose. The safeguard-120b weight is sized to fit on a single 80 GB GPU, while the smaller safeguard-20b targets lower-latency or constrained setups. Both sets are text-only reasoning models intended to run on user or host-controlled infrastructure.
Why this matters: until now, most teams either relied on a hosted API or built models from scratch. Open weights plus clear deployment guidance means organizations can run frontier reasoning and safety logic inside their own environments instead of sending data to someone else.
How it works, in plain terms: OpenAI handed out the model files and a permissive license, and it also supplied safety-focused weights trained to classify policy-related outputs. Think of it as getting both the engine and a set of built-in guardrails you can run where you choose.
What changes now: startups, enterprises, and hosting providers can experiment with and ship self-hosted reasoning and trust-and-safety tools. Practical limit: running these models still requires serious GPU hardware and engineering work, so this is a barrier lower, not gone.
What to watch next: will independent teams turn these open weights into reliable, production-ready services that compete with closed offerings, and how will customers and regulators respond as capable models become easier to self-host?