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, and then this morning switched to 27B, 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.

```
#!/bin/bash
llama-server \

-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

```

```
#!/bin/bash
export ANTHROPIC_AUTH_TOKEN="ollama"

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.

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.

Kicking off projects in the evening every now and then, that's a payback period of, what, maybe a couple months?

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