jesusluque/qwen3-30b-topiary-w576-code

Qwen3-30B-A3B — Topiary w576 (code calibration)

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:

pip install mlx-lm
mlx_lm.generate --model jesusluque/qwen3-30b-topiary-w576-code --prompt "..."

The equal-bytes head-to-head that motivated Topiary

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

Checkpoint	GB	Code PPL ↓	WikiText PPL ↓
This model (w576-fine, 4-bit)	13.10	2.87	12.98
Community 3-bit, full width	13.36	3.26	15.70
Mixed 3–4-bit	14.00	3.07	13.38

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.

Calibration

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

Provenance

• Base: Qwen/Qwen3-30B-A3B (Apache-2.0)
• Build: neuron-granularity permutation of the original bf16 experts, truncation to k=576 of 768, then 4-bit g64 quantization.
• All numbers reproducible from frozen configs; greedy decoding, fixed seeds.