jesusluque/qwen3-30b-topiary

Qwen3-30B-A3B — Topiary (fine, per-layer taper)

The flagship Topiary checkpoint: the routed experts of Qwen/Qwen3-30B-A3B, sculpted by routed activation salience at neuron granularity and truncated with a depth-tapered per-layer budget (576 neurons per expert in the shallowest layers rising to 704 in the deepest; mean 640 of 768), then quantized to 4-bit g64. No training, no distillation. 14.46 GB of weights.

The per-layer taper follows a measured law: sensitivity to width cuts increases monotonically with depth (late-layer error has no remaining layers to damp it). A controlled three-direction comparison at equal bytes showed the reverse allocation is the worst possible choice.

⚠️ Loading

Per-layer widths are not expressible in a standard mlx-lm config, so this checkpoint needs a ~30-line loader shim (src/per_layer.py in the Topiary repository) that patches block construction in layer order:

from per_layer import maybe_patch
from mlx_lm import load, generate

maybe_patch("jesusluque/qwen3-30b-topiary")
model, tokenizer = load("jesusluque/qwen3-30b-topiary")

If you need a zero-shim checkpoint, use the uniform sibling jesusluque/qwen3-30b-topiary-w640 (slightly lower scores).

Results (14.46 GB, Apple Silicon, MLX — all arms equal bytes)

Signal	This model	Uniform fine w640	Community 3-bit
Code PPL ↓	2.64	2.70	3.26
WikiText PPL ↓	10.27	10.35	15.7
GSM8K (50)	94%	94%	88%
MMLU (100, generative)	70%	74%	57%
HumanEval (50)	92%	84%	76%
HellaSwag (100)	70%	65%	66%
ARC-Challenge (100)	48%	49%	44%
Decode	76.6 tok/s	80.6	78.5

Honest notes: the per-layer taper trades a little broad-knowledge ranking (MMLU 70 vs 74, ARC −1) for large generation gains (HumanEval +8, both perplexities, HellaSwag +5). MATH-500 remains the 3-bit's stronghold (46% vs 40% here — distribution tails favor diffuse bit noise over salience truncation), while IFEval lands at 74% (3-bit: 68%). Pick the uniform sibling if MMLU-style ranking dominates your use.

Calibration & build

Mixed corpus (~40% code, ~30% GSM8K-train, ~30% WikiText), 151k tokens, routed-only statistics. Build: neuron-granularity permutation of the original bf16 experts, per-layer truncation (ratio 0.85 linear taper, floor 576), then 4-bit g64 quantization. All numbers reproducible from frozen configs; greedy decoding, fixed seeds.