Pi.dev, Qwen3.6-27B spark starter debate

Pi.dev looks capable, but it is deliberately spartan, which makes it a good fit for tinkerers and a rough on-ramp for beginners. If you want the shortest path to shipping code, a more batteries-included agent or a VS Code extension will probably teach the workflow faster before you drop into a CLI-first harness.

• –Pi.dev positions itself as a minimal terminal coding harness and explicitly skips features like sub-agents and plan mode, so the user has to build the workflow mentally instead of inheriting it.
• –Qwen3.6-27B is a legitimate local coding model choice; Qwen’s own docs say extended context matters and recommend keeping at least 128K context if you want thinking behavior to stay intact.
• –The reported 20-22 tok/s on a 2080 Ti is usable for agent loops, but agentic coding is usually bottlenecked by task decomposition, tool use, and iteration quality, not just raw generation speed.
• –If “thinking” is not showing up in Pi, the likely culprit is the client/template stack rather than the model being incapable; Qwen documents distinct thinking and non-thinking modes, but minimal harnesses do not always expose that switch cleanly.
• –For a first real workflow, VS Code plus an agent extension or a fuller CLI like OpenCode is usually less frustrating, then Pi.dev makes more sense once the user knows what knobs they actually need.