Qwen 3.5 27B hits Claude Code wall

This is less a Qwen problem than an agent-integration problem: Claude Code-style workflows are so prompt-hungry that 16K disappears fast once system prompts, tool traces, and requested output are all counted.

• –Qwen 3.5’s official docs support much larger contexts in compatible serving stacks, so the 16K cap is a deployment choice, not a hard model ceiling
• –The error shows 86,557 input characters plus a 16K output request, which leaves essentially no usable room on a 16,384-token backend
• –For coding agents, context compaction and lower `max_tokens` often matter as much as raw model size
• –If the goal is smooth Claude Code interoperability, a larger-context backend is the practical fix