jesusluque/qwen3-30b-topiary-w640

Qwen3-30B-A3B — Topiary w640 (mixed calibration)

A Topiary checkpoint: the routed experts of Qwen/Qwen3-30B-A3B were sculpted by routed activation salience — each expert's 768 intermediate neurons ranked by measured contribution on real traffic, permuted (a free, function-preserving transformation) and truncated to their top-640 prefix — then quantized to 4-bit (group size 64). No training, no distillation.

This is a standard mlx-lm checkpoint. It loads and serves with stock mlx-lm, no custom code:

pip install mlx-lm
mlx_lm.generate --model jesusluque/qwen3-30b-topiary-w640 --prompt "..."

Why

At equal memory, sculpting width beats lowering bits: quantization blurs every weight equally, while salience truncation concentrates the damage on what the model barely used. See the Topiary repository for the method, the full evaluation and the tools to build your own.

Results (14.46 GB of weights, Apple Silicon, MLX)

Signal	This model	Community 3-bit (13.4 GB)	Mixed 3–4-bit (14.0 GB)
Code PPL ↓	2.70	3.26	3.07
WikiText PPL ↓	10.35	15.7	13.4
GSM8K (50)	94%	88%	82%
MMLU (100, generative)	74%	57%	59%
HumanEval (50)	84%	76%	—
IFEval (50)	76%	68%	—
ARC-Challenge (100)	49%	44%	—
Decode	80.6 tok/s	78.5 tok/s	—

Honest limits — where the 3-bit keeps an edge: distribution tails. MATH-500 46% vs 38% here, HellaSwag 66% vs 65%. Bit reduction keeps every capability blurry-but-alive; width truncation sacrifices tails the calibration never exercised. Pick per use case.

Calibration

Mixed corpus (~40% code, ~30% GSM8K-train, ~30% WikiText), 151k tokens, routed-only statistics: each expert's salience is accumulated only over the tokens the router actually sends to it. The calibration corpus is the method's main dial — recalibrating on your own domain data takes minutes with the repository tools.

Provenance

• Base: Qwen/Qwen3-30B-A3B (Apache-2.0)
• Build: neuron-granularity permutation of the original bf16 experts, truncation to k=640 of 768 (−17% expert width), then 4-bit g64 quantization.
• All numbers reproducible from frozen configs; greedy decoding, fixed seeds.