Qwen 3.6 27B vs DeepSeek: local coding session

This uncut coding session tests Qwen 3.6 27B as a local LLM on a 128GB M3 Max Mac. The setup uses Q6 quantization with multi-token prediction, served through llama.cpp and driven by the Pi coding agent. The task is to build a React and Express app that produces nine scored YouTube titles and a thumbnail for a selected title.

The unusual requirement is that the app must call Codex exec through the speaker's subscription rather than call text or image APIs directly. Qwen handles the coding locally, but the finished app depends on Codex for generation. The session therefore does not establish an entirely offline workflow. The speaker supplies a handoff document from another agent explaining the commands.

The first generation fails. Server crashes, missing command flags and stalled requests lead to repeated debugging and user feedback. Qwen initially rewrites entire files, then later makes targeted edits. The speaker questions whether Pi causes the rewriting behavior, but the session does not settle that explanation.

By the end, the app produces title suggestions and a thumbnail. The speaker reports roughly 52 minutes, compared with two hours and a GPT hint in an earlier DeepSeek run. This is a single task comparison, with different prompting details, rather than evidence that Qwen can generally replace Claude Code.