Qwen/Qwen3.6-35B-A3B with 20% of its MoE experts structurally pruned with REAP (PRUNE method), produced with llmcompressor.
config.json records the new expert count (num_experts: 205); expert
indices are remapped contiguouslymtp.layers.0) was not calibrated and is left untouchedREAP scores each expert by the router-weighted mean of its output activation
norms over the calibration tokens routed to it,
S_j = mean(g_j * ||f_j||_2), using forward computation only (no gradients or
recovery training). This model was calibrated on all 2,048 samples of the
RazorCal
multi-domain calibration corpus; see the
REAP expert pruning example
for the calibration and pruning scripts.
from transformers import AutoProcessor, Qwen3_5MoeForConditionalGeneration
model = Qwen3_5MoeForConditionalGeneration.from_pretrained(
"inference-optimization/Qwen3.6-35B-A3B-REAP20", torch_dtype="bfloat16"
)
processor = AutoProcessor.from_pretrained(
"inference-optimization/Qwen3.6-35B-A3B-REAP20"
)
model-00001-of-00026.safetensors 3.5 GBmodel-00025-of-00026.safetensors 3.3 GBmodel-00006-of-00026.safetensors 3.2 GBmodel-00016-of-00026.safetensors 3.2 GBmodel-00008-of-00026.safetensors 3.2 GBmodel-00010-of-00026.safetensors 3.2 GBmodel-00018-of-00026.safetensors 3.2 GBmodel-00014-of-00026.safetensors 2.8 GBmodel-00012-of-00026.safetensors 2.8 GBmodel-00020-of-00026.safetensors 2.8 GBmodel-00022-of-00026.safetensors 2.7 GBmodel-00024-of-00026.safetensors 2.7 GBmodel-00004-of-00026.safetensors 2.7 GBmodel-00023-of-00026.safetensors 2.7 GBmodel-00003-of-00026.safetensors 2.7 GBmodel-00005-of-00026.safetensors 2.7 GBmodel-00026-of-00026.safetensors 2.1 GBmodel-00015-of-00026.safetensors 1.3 GBmodel-00021-of-00026.safetensors 1.3 GBmodel-00013-of-00026.safetensors 1.3 GBmodel-00002-of-00026.safetensors 1.2 GBmodel-00007-of-00026.safetensors 882 MBmodel-00009-of-00026.safetensors 882 MBmodel-00019-of-00026.safetensors 882 MBmodel-00011-of-00026.safetensors 882 MBmodel-00017-of-00026.safetensors 882 MBQwen/Qwen3.6-35B-A3B