Same Blender agent prompt, two models: mystery QoDuck builds a 1942 truck, Qwen builds a toy
Character-Apple-8471 · reddit · 2026-11-09
The author ran one 1,600-word prompt turning models into autonomous Blender 3D artists (plan, build in chunks, render, inspect, fix, checkpoint, continue) targeting a historically accurate 1942 German military truck.
Results
- QoDuck (free mystery model in Qoder, possibly a Qwen 3.8 variant) delivered a recognizable weathered Opel Blitz: crowned fenders, louvred hood, 15-slat grille, split windscreen, full interior, six lug nuts per wheel, canvas tilt, 14 final renders, all-procedural materials.
- Qwen3.8 Flash Next (Unsloth UD-Q4 GGUF) shipped a beige blockout: flat plank fenders, undersized wheels, floating running boards, untextured bumper.
Key insight: Qwen rendered nearly 3x more frames (150 vs 47) yet still had visible geometry errors — it kept the render loop but stopped closing it. Agentic 3D work hinges on whether the model acts on visual feedback, not on having the loop structure.
More from coding & agent
- AI Agents Waste ~18% of Output Tokens, Paper Finds; Pruning 41-46% Barely Hurts Success — rohanpaul_ai · 2026-11-09
- Lovable Is Opus 5.5 at API Prices: Claude Plan Does Same Work ~45x Cheaper — PawelHuryn · 2026-11-09
- Systems Engineer: Write Code Yourself in the AI Era to Build Ownership and Taste — Abhishekcur · 2026-11-09
- Prefixwise names why your prompt cache misses — 19 causes, recovers 99.9% of savings — No-Individual159 · 2026-11-09
- Dev spends 12.5 months of nights building a local, Claude Code-style meeting scribe — Jazzlike-Swing-6198 · 2026-11-09
- After a 100% AI coding mandate, codebase triples and nobody understands it — TechPreacher · 2026-11-09