Running OpenAI's GPT-OSS locally: the good, the bad, and the loopy

OpenAI recently launched something I didn’t expect to see from them: an open source model called GPT-OSS. After years of keeping their models locked behind APIs, they’ve finally released both 20 billion and 120 billion parameter models that you can run entirely on your own hardware. Both can run on consumer grade hardware, but realistically most people would go with the 20 billion model since it fully fits into 16 gigs of VRAM on a GPU, which is why I went with that one as well.