this weekend i messed around with codex some more to try something ... it found a lot to do and to look through and it kept taking an hour max until the 5h ratelimit thing was reached. this is a very annoying limitation. i guess i should see how well i can run some local qwen model or something got kinda curious now but it could be a tight fit on my hardware. i cant justify to throw money at this problem
0 Boosts
0 Likes
@milan
depending on how much work you want to throw at it; i run qwen3.8 27b and qwn3.6 35b-a3b both in UD-Q3_K_XL quant with 114668 q8_0 with mtp on two rtx 3060 - so if you have similar hardware laying around, it's sufficient for lighter tasks in a single user setup
it won't work in long agentic runs/tasks and you need to watch what it does, but for me, this setup, despite being rather slow with the dense model, has already saved so much time, i don't care how limited it is
the MoE is fast but dumb and the dense model is slow but very accurate, that's why i alternate between them depending on the task it should work on
my 2 cents