Qwen3.6-27B quants hold at IQ4_XS

Useful stress test, but it is still a narrow eval: it mixes board-state tracking, SVG layout, and visual correctness, so the ranking is directional rather than universal. The main takeaway is that Qwen3.6-27B degrades gracefully until the lowest quants, where spatial consistency starts to break.

• –Q8_0 and Q5_K_XL are the safest bets if you want near-BF16 behavior without paying full-precision memory costs
• –Q6_K is where the first visible errors show up, especially in piece placement
• –IQ4_XS appears to be the lowest quant that still feels usable on this task for a 16 GB card
• –IQ3_XXS mostly preserves piece state but can flip board orientation, which is fatal for rendering correctness
• –KV cache quantization and TurboQuant-style throughput gains matter almost as much as weight quant when local speed is the goal