Ran Q8_0, Google Q4_K_M, and Unsloth UD-Q4_K_XL through identical 8-question quality suites on the same hardware. Q8 (4.69 GB): 7/8 at 30.12 tok/s. Google Q4_K_M (3.21 GB): 6/8 at 35.36 tok/s. Unsloth UD (2.94 GB): 7/8 at 38.16 tok/s. The smallest model is the fastest AND ties the biggest for accuracy. Unsloth's dynamic quantization preserves the weights that matter and compresses the ones that don't — and the imatrix calibration data means it knows which is which. This isn't theoretical. 2.94 GB, 38 tok/s, 88% accuracy. Unsloth UD-Q4_K_XL is the daily driver. Full stop.