A 50%-expert-pruned MLX 8-bit derivative of mlx-community/Qwen3.6-35B-A3B-8bit using a per-layer adaptive budget (variable expert count per layer).

50% removed overall, but the per-layer count varies (109–186, mean 128) — more experts kept in language-structure-heavy early layers, fewer in language-specialized late layers (allocation by English routing warmth, floor 16). Cold (un-healed). At 50% the adaptive allocation gives a small perplexity edge over uniform (wikitext 10.93 vs 11.11) but is a wash on tasks.


Routed experts	109–186/layer (mean 128, −50% total)
Active	top-8 + 1 shared
Healing	none (cold)
Format	MLX, 8-bit

⚠️ Loading — variable expert counts require `strict=False`

Standard mlx_lm.load is strict and will fail with a shape mismatch. Load with strict=False:

python

from pathlib import Path
from huggingface_hub import snapshot_download
from mlx_lm.utils import load_model, load_tokenizer

path = Path(snapshot_download("neopolita/Qwen3.6-35B-A3B-8bit-Prune-50-Adaptive"))
model, _ = load_model(path, strict=False)   # adopts the per-layer shapes
tok = load_tokenizer(path)

The MoE forward routes on each layer's gate width + top-k (never a global num_experts), so the variable model runs correctly once loaded with strict=False.

Benchmarks (vs unpruned base; 100–200 samples)

metric	base	this model
wikitext PPL	7.98	10.93
winogrande	0.70	0.65
hellaswag (acc_norm)	0.555	0.58
piqa (acc_norm)	0.84	0.81
gsm8k (strict)	0.99	0.95
ifeval (prompt-strict)	0.86	0.76
ifeval (inst-loose)	0.933	0.847

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

neopolita/Qwen3.6-35B-A3B-8bit-Prune-50-Adaptive

Qwen3.6-35B-A3B-8bit-Prune-50-Adaptive

⚠️ Loading — variable expert counts require strict=False

Benchmarks (vs unpruned base; 100–200 samples)

neopolita/
Qwen3.6-35B-A3B-8bit-Prune-50-Adaptive

⚠️ Loading — variable expert counts require `strict=False`