Reddit Guide: Taming Qwen3.8 Overthinking with Task-Scaled Reasoning Budgets
Express_Quail_1493 · reddit · 2026-09-08
A Reddit user shares practical budgets to stop Qwen3.8 overthinking: 512–1,024 tokens for simple scripts, 2,048–4,096 for standard coding and refactoring, 8,192 for AIME-level math and nested debugging, 16,384 for competition-level tasks. Fine-tunes promising less thinking often degrade the model. Setting a --reasoning-budget-message of "ok now" teaches the pattern universally. At 2,048, Qwen3.8-27b solved an issue Gemini failed.
More from coding & agent
- Agent verified 'success' after double-voting: a lesson in harness vs model bugs — TejasKumar_ · 2026-10-26
- diginomica: agentic automation hype is overdone, and 'The Hackening' of enterprise security has begun — jonerp · 2026-10-26
- 'Vibe verification': ~60% of AI-written formal specs prove the wrong property — basedjensen · 2026-10-26
- Devin launches Security suite powered by Claude Mythos with $100M+ credits — DevinAI · 2026-10-26
- HackBench injects reward hacks into agent trajectories to stress-test monitors — DanielKhashabi · 2026-10-26
- Hugging Face ships hf CLI updates: deduped cache saves 19GB, faster buckets, retryable GPU jobs — yb2698 · 2026-10-26