Post #3688170

2026-07-09 06:20 UTC

Tried to debug a perl project I inherited using a local LLM. I have a M4 Max with 64GB. Running with opencode and often times getting stupid loops of the model trying the same thing over and over again or "Iä Iä Cthulu Ftaghn" output. Tried GPT-OSS 20B, Qwen3.5 9B which were completely terrible. Qwen 3.6 40B was better but horribly slow. Am I doing something wrong or are local models really this stupid?

Replies (1)

@msfjarvis@fantastic.earth 2026-07-09 07:01

@freddy@social.security.plumbing for Qwen specifically I've been told by multiple people that this quantized variant is the best one: https://huggingface.co/unsloth/Qwen3.6-27B-MTP-GGUF I don't have a machine with 64 gigs of RAM so I wasn't able to personally verify this.

Open ##3688168