Lately, I’ve been experimenting more and more with open-weight models and I think GPT-OSS-120B just nails it: it’s fast, powerful and it listens quite good the the instructions it’s given.

For some context: I’m working on an AI Assistant at one of the companies I work for and we sometimes have some content that does not fully aligns with content policies of proprietary model providers such as alternative healing for example.

Models like GPT-5.2 Chat do not answer questions when asked about such topics, even if given the instruction to do so based on the document we pass the model. This is very annoying because some people do want to try alternative healing methods and a model refusing because it does not fully align is in my opinion very bad.

This is where GPT-OSS-120B shines: it listens very good to the instructions you give it. And no, you cannot use it for the most extreme forms of content, which is absolutely fair and I am happy that it protects users. But a model that does what you ask it to do: great!

Personally I also like using GPT-OSS-120B for fast Q&A and paired with Brave’s AI Search API for real time answering in OpenWebUI makes it feels almost as snappy as Perplexity (which is also a great tool!).