Qwen3.8-27B quant keeps 99% of reasoning at 15% size

A task-tuned quant holds Qwen3.8-27B at 82.81% held-out reasoning accuracy, just under BF16's 83.59%, at 15% of the original size, per a TAK post to r/LocalLLaMA.

Against a byte-matched Unsloth UD IQ2_S quant at the same size, TAK's accuracy edge is 5.47 points at 27B. That gap widens as models shrink: +11.72 points on Qwen3.5-4B, +14.06 on Gemma 4 E4B, +19.53 on Gemma 3 4B QAT.

The method builds an importance matrix from a corpus matched to the target task, then finds the smallest size before the model's output collapses. From there it trades precision between tensors inside a fixed byte budget, no pruning, no fine-tuning, no merging. It moves bits to where the calibration data says they're needed.

The author added a warning after the post went up. The reasoning-specialized quant breaks down on coding, where testers hit repetition loops. That's a real limit. The quant is task-specific, not a general swap-in for full precision.