AI chessboard test: Qwen3.6 and ZAYA1 models battle SVG generation accuracy

A Reddit user conducted a quality test comparing several local and cloud-based AI models on their ability to generate a chessboard SVG. The test involved Qwen3.6 variants (27B and 35B-A3B) in different MLX quantizations, the open-weight ZAYA1 8B, and derivatives like GRM 2.6 Plus from OrionLLM. Results show that higher bit counts don't guarantee better performance: Qwen3.6 27B MLX oQ6 (6-bit) delivered good, correct output but lacked row/column labels, while its 3.5-bit oQ3.5e version was poor. Surprisingly, Qwen3.6-27B-neo-code-di-imatrix-max at 4-bit (iq4_nl) performed well, but its 5-bit (q5k_s) variant was totally wrong.

ZAYA1 8B demonstrated perfect SVG generation when accessed via the Zaya cloud playground (presumably FP16), but local inference using MLX-LM failed due to the model's reliance on the RSA technique—the 8-bit quant entered a reasoning loop without producing output. This suggests that local inference engines need better support for RSA-based models. The test also explored derivative models like GRM 2.6 Plus and Qwopus 35B-A3B-v1, with mixed results; GRM 2.6 Plus Q4K_M (4-bit) was correct and visually good, but its 3-bit version degraded. Overall, the smallest viable quant was a 27B model at 3-bit (Q3K_M), hinting at potential for lightweight deployment on consumer hardware (<12GB RAM).