After several failed attempts, I've finally managed to get a viable Opencode loop working with a local model: Qwen3.8-27b (Q4_K_M).
Tool calling is much more stable than configuration that I tested before (OK, Ling-3.0-tiny, one of the more recent Chinese models with 8b parameters, was also quite stable).
This is certainly less performant than commercial services, e.g., Copilot with a model like Gemini, but pushing the limit is part of the fun, isn't it?
I've upgraded my inference machine (11th gen Core i7, 32GB RAM, fast storage), installed Incus (Zabbly stable) and use a Debian 13 container as the Opencode sandbox; llama-swap has learned a few new tricks, too. My two RTX 3060 are humming along, crunching numbers.
#homelab #Qwen38 #rtx3060