50% of routed experts (256 → 128/layer, uniform) removed by English-cold frequency selection. This is the cold (un-healed) prune: at 50% compression, cheap router-only healing recovers perplexity but degrades instruction-following, so the un-healed model is the stronger downstream checkpoint (genuine recovery at this ratio needs heavy expert-level distillation).


Routed experts	256 → 128/layer (−50%), uniform
Active	top-8 + 1 shared
Healing	none (cold)
Format	MLX, 8-bit

Benchmarks (vs unpruned base; 100–200 samples)

metric	base	this model
wikitext PPL	7.98	11.11
winogrande	0.70	0.65
hellaswag (acc_norm)	0.555	0.57
piqa (acc_norm)	0.84	0.80
gsm8k (strict)	0.99	0.97
ifeval (prompt-strict)	0.86	0.74
ifeval (inst-loose)	0.933	0.865

At 50% the prune has a real cost (instruction-following hit hardest). Use the 30% variants for near-lossless compression.

Usage

python

from mlx_lm import load, generate
model, tok = load("neopolita/Qwen3.6-35B-A3B-8bit-Prune-50-Uniform")
print(generate(model, tok, prompt="Hello", max_tokens=128))

Research artifact. Selection by per-layer English routing frequency; small-sample benchmarks (100–200). License/usage inherit from the base model.