Qwen 3.6 Built a Full App on My Mac in One Uncut Session

I recorded a full session of Qwen 3.6 building a complete full-stack app on my Mac, no cloud API, no help from Claude or ChatGPT, in about 52 minutes. The video is the uncut run. This post is the short version, and the setup guide is below it, because the setup is the part you actually reuse.

What happened in the build

The model is Qwen 3.6 27B at Q6 quantization with multi-token prediction, served through llama.cpp and driven by the Pi coding agent, on a 128GB M3 Max. I picked Q6 over Q4 or Q8 for a bit more quality with some speed, and RAM and CPU stay on screen the whole run because people asked for that last time. The rule I set makes it harder: the app cannot call any cloud API directly, so to make its titles and thumbnails it has to drive codex exec, a handoff the model has almost certainly never seen. In the video the first generation fails, the server crashes, and Qwen debugs it and finishes anyway. For reference, DeepSeek took two hours on the same test and needed a hint from GPT. Qwen did it in under an hour on its own. Watch the video for the full play-by-play.

How to set it up yourself

I made a separate setup video so you can get this running without watching the whole build. Here is everything in it.

Alibaba released Qwen 3.6, and this walkthrough gets it running locally as a Claude Code alternative. You use Ollama to run the model and Qwen Code, an open-source agent built like Claude Code, to drive it. The steps are install Ollama and pull Qwen 3.6, install Qwen Code, then wire it to the local model through settings.json: base URL pointing at localhost:11434, the exact model ID, and the API key set to the word Ollama since the SDK needs something even though Ollama does not check it. The number-one gotcha is context length. Ollama defaults to 4096 tokens, so the agent looks like it is forgetting things mid-task when it is really truncating silently. Raise it in the settings and on the Ollama serve side. One more thing that trips people up: Qwen Code says OpenAI even when it is fully local, because it uses the OpenAI-compatible SDK format, not OpenAI itself. From there I confirm it can read and write files, then run three demos against Claude Code: a 3D aquarium, a designer portfolio, and a macOS desktop clone. Qwen matched Claude on the aquarium fish, ran close on the portfolio, and got the look of the desktop but left the icons non-clickable.

Both videos above show the setup and the uncut build. If you want the cost comparison between running local and a paid plan for your own workload, the calculator on openclawdc.com will run it.

I specialize in standing up local coding agents that hold up on real work, not just demos. Book a call at cloudyeti.io/meet.