Interesting, this is still gpt-oss-20b, no jailbreak or anything, just about 50k tokens of Monster Girl Encyclopedia I (without monster cards) in the description, "developer" prompt, after I asked it to explain what Sabbaths are.

On a related note, long context doesn't really take a lot of VRAM, but due to the sliding window it reprocesses the prompt every time by default (in llama.cpp), and for some reason prompt processing seems much slower than it should be, even after setting batch size to 8k tokens.