Qwen3.6-35B-A3B · Winnow keep-50
Qwen/Qwen3.6-35B-A3B with 50% of routed-expert
channels removed by winnow variable-width expert pruning —
18.5B parameters. Requires trust_remote_code=True (ragged variable-width experts,
PrunedQwen3_5MoeForCausalLM); transformers ≥ 5.
Instead of deleting whole experts (REAP-style), winnow shrinks each expert's SwiGLU intermediate
width according to a per-channel importance score (router weight × |post-SwiGLU activation| ×
down-projection column L2), allocated by a per-layer global rank in blocks of 128 channels.
Attention, embeddings and the router are untouched. Calibration: C4, qwen35_stats1024.
Results (unhealed, this checkpoint)
| Keep fraction | 0.50 |
| Params | 18.5B (base 35B) |
C4 ppl, 32 × 2048-token seqs (winnow_verify.json) | 14.69 |
Whole-expert REAP comparison at matched budget and calibration, and downstream numbers (MMLU-Pro / IFEval, Protocol E), are in the Winnow writeup. Healed variants: keep50-healed, keep25-healed.
Serving
vLLM via the winnow plugin in
variable-reap; --enforce-eager, pipeline-parallel
supported (keep-50 runs on 2 × 24 GB).