A Topiary checkpoint: the routed experts of Qwen/Qwen3-30B-A3B sculpted by routed activation salience, truncated to their top-576 of 768 intermediate neurons (−25% expert width), then quantized to 4-bit (group size 64). Calibrated on code — this is the coding-slanted sibling of the mixed-calibration w640; smaller, sharper on code, weaker on math.

Standard mlx-lm checkpoint — loads with stock mlx-lm, no custom code:

At the same memory budget, against the community 3-bit quantization:

Calibrated on code, it still wins general text by −17% PPL. Other signals: HumanEval 84%, GSM8K 78%, MMLU 64%, ARC 43%, HellaSwag 60%, ~55 tok/s decode.

Honest limit: the code-only calibration trades math away (GSM8K 78% vs 94% of the mixed-calibration w640 sibling). The calibration corpus is an amplitude-vs-specialization dial — choose the sibling that matches your use.

Code corpus (The Stack: Python + C++ + Swift), routed-only statistics (salience accumulated only over tokens the router sends to each expert).

The taper flagship of this family was compared against the UNPRUNED Qwen3-30B-A3B-4bit on a four-benchmark suite (n=500 on MMLU/LAMBADA): reasoning intact (MATH-500/MBPP), but -10 MMLU / -4.4 LAMBADA points — salience pruning removes knowledge, not reasoning. This model shares the pruning method and was NOT re-measured on that suite; expect a knowledge cost of similar nature. If broad world knowledge dominates your workload, prefer the unpruned checkpoint. Details: github.com/jesusluque/topiary-stream (paper §2.4).