The author evaluated multiple quantizations of the Qwen 3.8 27B vision model—bf16, Q4_K_M, Q3_K_M, a custom 12 GB quant, and Ternary Bonsai 2—by giving each a single PNG brief and the sentence “Build what the at” to generate an animation, producing side‑by‑side videos and per‑check scores. This follow‑up addresses prior criticism that Wikitext perplexity alone does not demonstrate quantization usefulness, showing that the custom 12 GB quant matches the Q4_K_M perplexity while reducing file size by ~27%.