Qwen3.5 local coding benchmark disappoints

This is more an orchestration problem than a model verdict. For local coding, the agent loop, prompt shape, and sampling defaults matter almost as much as the model family.

• –Qwen3.5's docs recommend conservative coding settings like `temperature=0.6`, `top_p=0.95`, `top_k=20`, and `presence_penalty=0.0`; a generic chat preset can make the same model look far worse than it is.
• –The official Qwen3.5 cards put the dense 27B slightly ahead of the 35B-A3B on SWE-bench Verified, so bigger total parameter counts are not automatically better for coding.
• –Context length helps the model remember repo state, not magically improve reasoning; once the thread balloons, you pay more latency and lose focus.
• –Claude Code can route to local backends, but small models need shorter, test-driven tasks and tighter prompts to stay on rails.
• –If you want a fair benchmark, use edit-run-fix loops with pass/fail tests instead of a one-shot “build me a game” prompt.