Qwen3.8-27B as local coding agent burns 16k tokens thinking, never calls tools
previaegg · reddit · 2026-11-06
A Reddit user running Qwen3.8-27B (UD-Q5KXL) locally on consumer hardware (RTX 4080 Super + 4070 Ti Super, llama.cpp + OpenCode) reports severe "excessive thinking" in agentic coding. Inference speed is fine (2,330 prefill / 24.1 t/s).
The experiment: comparing medium vs low reasoning effort on the same task, both runs made only 2 inspection tool calls, zero implementation calls, generated 16,000+ output tokens exhausting the limit, failed the task, and spent 12 minutes. Lowering reasoning effort didn't help.
Open questions (post includes the full llama-server launch command: 64K ctx, q80 KV cache, flash-attn):
- Are --reasoning/--reasoning-effort/--reasoning-preserve right for agentic coding, or should reasoning be disabled?
- Could OpenCode's system prompts/tool definitions be the cause? Would another coding agent work better?
- Would vLLM/SGLang improve tool-calling reliability? Does Q5KXL quantization hurt instruction following?
- Context management strategies, and whether headless Linux helps.
The thread gathers practical experience from people running local models as coding agents.
More from coding & agent
- Musecases now catalogs 600+ real-world Meta Muse AI agent use cases — thisiskp_ · 2026-11-06
- Claude Code can now edit videos — first preview in under 20 minutes — PrajwalTomar_ · 2026-11-06
- Siraj Raval: better plumbing beats a bigger model in his code-free VR museum build — sirajraval · 2026-11-06
- Documentation Is Part of Done: Put It in Your Definition of Done — dSebastien · 2026-11-06
- Coding agent overhead was ~90% of spend; task-based attribution cut it to 45% in 5 weeks — entelligenceai17 · 2026-11-06
- Viktor gives its AI employee an email address — just CC it to delegate work — hey_abusiddik · 2026-11-06