Task	Metric	Value
arc_challenge	sample_len	1172.0000
arc_challenge	acc,none	0.5256
arc_challenge	acc_stderr,none	0.0146
arc_challenge	acc_norm,none	0.5614
arc_challenge	acc_norm_stderr,none	0.0145
arc_easy	sample_len	2376.0000
arc_easy	acc,none	0.7891
arc_easy	acc_stderr,none	0.0084
arc_easy	acc_norm,none	0.7929
arc_easy	acc_norm_stderr,none	0.0083
winogrande	sample_len	1267.0000
winogrande	acc,none	0.7174
winogrande	acc_stderr,none	0.0127
piqa	sample_len	1838.0000
piqa	acc,none	0.7807
piqa	acc_stderr,none	0.0097
piqa	acc_norm,none	0.7933
piqa	acc_norm_stderr,none	0.0094
hellaswag	sample_len	10042.0000
hellaswag	acc,none	0.5632
hellaswag	acc_stderr,none	0.0049
hellaswag	acc_norm,none	0.7403
hellaswag	acc_norm_stderr,none	0.0044
openbookqa	sample_len	500.0000
openbookqa	acc,none	0.3440
openbookqa	acc_stderr,none	0.0213
openbookqa	acc_norm,none	0.4600
openbookqa	acc_norm_stderr,none	0.0223

Usage

Method: mobe
Base model: Qwen/Qwen3-30B-A3B
Compressed: gate_proj, up_proj over layers 0-48 ( down_proj kept dense)
Settings: mobe, num_B=32, activation=tanh

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "AverageMetaheuristicsEnjoyer/qwen3-30b-a3b-mobe-b32-h100",
    trust_remote_code=True,
    torch_dtype="bfloat16",
    device_map="auto",
)
tok = AutoTokenizer.from_pretrained("AverageMetaheuristicsEnjoyer/qwen3-30b-a3b-mobe-b32-h100", trust_remote_code=True)

The factored experts are reconstructed at runtime via the bundled Qwen3MoBEForCausalLM.