[kstenerud]

Opencode can be connected to Qwen for coding grunt work. It's pretty decent, but I prefer to use Claude. Less mistakes, better for reasoning architectural decisions with the agent checking against the existing codebase and documentation. I also keep things sandboxed in yoloAI for peace of mind and to eliminate permission fatigue.

[joelthelion]

Do you pay for the model? Or run it locally? If so, which hardware do you use? Is it fast?

[kstenerud]

I use an RTX 3090 with 24gb on it, which can comfortably fit https://huggingface.co/Qwen/Qwen3.6-27B. It's freely downloadable, so you only pay for the gfx card and the electricity. There are also plenty of models out there that work well on Apple unified memory.

But mostly I just use Claude now since I already have a max subscription. I've done hybrid setups in the past where I had Claude do the hard stuff and a local LLM do the code monkey stuff, but with the right optimizations I'm no longer hitting the Claude token ceiling.

[joelthelion]

Thank you !