Am I Expecting Too Much?
这条记录涉及编程工具或代码能力更新，适合开发者评估工作流变化和可复用价值。
I’ve been trying to get started with local coding help after finding Claude Code really useful but I’m just running into a ton of problems.
On my RTX 5090, I’ve been running Qwen 3.6 27B UD_4 at 131K context, no KV quants, quant’d by Unsloth. I’ve been using Cline in VS Code, as it seemed the closest parallel to Claude Code.
Recently, I wanted to try implementing some plans I had made with Fable. I had Fable prepare detailed, junior-engineer-friendly implementation plans for a pretty basic Python app, which I threw into the repo.
But I just can’t get anywhere with Qwen. It writes code that contains a ton of mistakes, tries to write terminal commands that are just broken with basic syntax errors, etc. It can’t follow even these very detailed plans going basically step by step.
To be clear, I’m not expecting Opus level performance or even Sonnet level work, but it just doesn’t work at all, even within the harness and via basic terminal calls.
Is something in my approach or setup wrong?
Is there a better harness or model I should swap to?