Local test: Qwen 3.8 27B's overthinking brings it near Sonnet-level performance
maxwell321 · reddit · 2026-08-17
A user running local models on 3x 3090s plus a Tesla P40 reviewed Qwen 3.8 27B (Unsloth UD-Q8KXL quant), concluding its heavy thinking draws out fine details from training knowledge, landing near Claude Sonnet with potential for Opus-level results.
The benchmark: 1:1 recreations of classic arcade games. Qwen 3.6's Galaga clone was effectively a Space Invaders clone — no diving enemies, no shooting, missing details requiring hand-holding. Qwen 3.8 built sprites from dynamic pixel bitmaps with two-frame animation, added a CRT filter and power-on simulation, and even remembered the fighter-capture mechanic (rescue for dual ships), only replacing the capture beam with collision. The author says this closes much of the local-vs-proprietary gap, worth the extra inference time and context.
More from Models
- Prediction: frontier AI revenue growth could slow even as AI usage keeps rising — yi111 · 2026-10-16
- Meta teased 816B-param open-weight Muse Spark Open with 1M context, says meme thread — alexandr_wang · 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