 | Launched claude code, pointed it at my running Qwen, and, well, it vibe codes perfectly fine. I started a project with Qwen3.6-35B-A3B (Q4) yesterday, and then this morning switched to 27B (Q8), and both worked fine! Running on a dual 3090 rig with 200k context. Running Unsloth Q_8. No fancy setup, just followed unsloths quickstart guide and set the context higher. ``` -hf unsloth/Qwen3.6-27B-GGUF:Q8_0 \ --alias "unsloth/Qwen3.6-27B" \ --temp 0.6 \ --top-p 0.95 \ --top-k 20 \ --min-p 0.00 \ --ctx-size 200000 \ --port 8001 \ --host 0.0.0.0 ``` ``` export ANTHROPIC_API_KEY="" export ANTHROPIC_BASE_URL="http://192.168.18.4:8001" claude $@ ``` The best part is seeing Claude Code's cost estimate. Over that 8 hours I would have racked up $142 in API calls, and instead if cost me <$4 in electricity (assuming my rig pulled 1kw the entire time, in reality it's less, but I don't have my power meter hooked up currently). So to all the naysayers about "local isn't worth it", this rig cost me ~$4500 to build (NZD), and thus has a payback period of ~260 hours of using it instead of Anthropic's API's. If I use it full time as my day job, that's ~30 days. If I run a dark-software factory 24/7, that's 10 days.Kicking off projects in the evening every now and then, that's a payback period of, what, maybe a couple months? What did I vibe code? Nothing too fancy. A server in rust that monitors my server's resources, and exposes it to a web dashboard with SSE. Full stack development, end to end, all done with a local model. I interacted with it maybe 5 times. Once to prompt it, and the other 4 for UI/UX changes/bug reports. I'm probably not going to cancel my codex subscription quite yet (I couldn't get codex working with llama-server?), but it may not be long [link] [comments] |