For me it just returns a wall of text with (part of) its reasoning about composing a short roast, but not the roast itself. I guess this is not intended.
This ended up impacting quite a few users; the response is streamed from a Vercel function where I had not changed the default timeout, so at 10 seconds it was just cutting things off where it was. Some invocations of the DeepSeek API take much longer than others. This should be fixed now with a longer timeout.
API keeps going through periods of 1-2 minutes of everything timing out... then works well for ~5-10 minutes. Might be better to host this model yourself for serious use right now.
