Qwen3.6-35B-A3B flunks Godot camera planning

The hot take: this looks less like “Qwen is dumb” and more like a reminder that local model quality is brittle once you combine sparse MoE, heavy quantization, and a niche planning prompt.

• –The Q4/IQ4 quantization choice could easily wipe out enough structure-following quality to make the output look random.
• –Godot scene-tree planning is a domain-specific software architecture task, not a generic chat benchmark, so instruction tuning matters a lot.
• –Gemma 4 may simply be better aligned to this exact style of “plan only, no code” prompt.
• –A single anecdotal prompt is not a reliable model ranking; repeatability across prompts and quant levels matters more.
• –If Qwen3.6 also skipped SpringArm3D, that suggests a failure in procedural decomposition, not just missing trivia.