Qwen3.8 27B xhigh burns 23.8K tokens overthinking a simple card prompt — and crushes every rival
hiImMate · reddit · 2026-08-18
A user tested Qwen3.8 27B (UDQ4XL quant) thinking-effort tiers against DS V4 Flash, ChatGPT, and Claude Opus 5, using an open-ended prompt to write an HTML card about apple benefits.
- xhigh spent 23.8K tokens thinking and produced a result far above everything else; medium thought only 794 tokens and output was mediocre.
- DS V4 Flash used 927 tokens; ChatGPT and Claude thought lightly and also trailed.
- With a detailed prompt specifying design specifics (orchard-note style, perforation, stats section), medium produced 3.3K tokens and a result close to xhigh quality.
Takeaway: xhigh can get into a long thinking match even on simple prompts but delivers notably better output; with concrete prompts, medium cuts tokens/time heavily while staying strong — good for low-tps setups.
More from Models
- Prediction: frontier AI revenue growth could slow even as AI usage keeps rising — yi111 · 2026-10-16
- Nolla Health Post-Trains Nemotron 3.5 VL: Visual Diagnosis Jumps from 35% to 66.8% — LoganGrasby · 2026-10-16
- Dev argues coding models now beat nearly all engineers in every branch of software — zack_overflow · 2026-10-16
- Prediction: Chinese frontier models could match GPT-6 by March 2027 — teortaxesTex · 2026-10-16
- Liquid AI's d1-ultra, #1 on Artificial Analysis Decision Index, lands on Vercel — JosephJacks_ · 2026-10-16
- Weekly AI roundup: Claude Haiku 5.5 cuts prices ~90%, Gemini 4 Argon output jumps to 1M tokens — DeepLearningAI · 2026-10-16