Downloads · 30 days
22
8% of all-time downloads
neopolita/Qwen3.6-35B-A3B-8bit-Prune-30-Adaptive
Qwen3.6-35B-A3B-8bit-Prune-30-Adaptive is a machine learning model from neopolita. Use it for the machine learning task on the model card, and read the license before you ship it in a product. It is set up for mlx. The card lists the license as apache-2.0.
A 30%-expert-pruned and router-healed MLX 8-bit derivative of mlx-community/Qwen3.6-35B-A3B-8bit, using a per-layer adaptive budget (variable expert count per layer).
Downloads · 30 days
22
8% of all-time downloads
All-time downloads
260
Public
Parameters
25B
26.5 GB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors26.5 GB · 100%
How the weights are stored.
U3225B · 100%
From the Hugging Face model README
A 30%-expert-pruned and router-healed MLX 8-bit derivative of mlx-community/Qwen3.6-35B-A3B-8bit, using a per-layer adaptive budget (variable expert count per layer).
30% removed overall, with per-layer counts varying (143–245, mean 179) — more experts kept in language-structure-heavy early layers, fewer in language-specialized late layers (allocation by English routing warmth, floor 16). Routers then recalibrated with Router-KD (rank-64 LoRA on mlp.gate, ~5.7M params, wikitext calibration), fused.
| Routed experts | 143–245/layer (mean 179, −30% total) |
| Active | top-8 + 1 shared |
| Healing | Router-KD (router-only) |
| Format | MLX, 8-bit |
strict=FalseStandard mlx_lm.load is strict and will fail with a shape mismatch. Load with strict=False:
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-30-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.
| metric | base | this model |
|---|---|---|
| wikitext PPL | 7.98 | 8.63 |
| winogrande | 0.70 | 0.725 |
| hellaswag (acc_norm) | 0.555 | 0.585 |
| piqa (acc_norm) | 0.84 | 0.81 |
| gsm8k (strict) | 0.99 | 0.97 |
| ifeval (prompt-strict) | 0.86 | 0.82 |
| ifeval (inst-loose) | 0.933 | 0.896 |
At 30% the prune is near-free on most downstream tasks. (On the uniform 30% variant, Router-KD also fully recovers instruction-following / IFEval; on this adaptive variant the heal is roughly task-neutral and mainly recovers perplexity.)
Research artifact. Selection by per-layer English routing frequency (not REAP); small-sample benchmarks (100–200). License/usage inherit from the base model.