List your company

Language Modeling on OpenWebText (held-out set)

11.5PPLGrMoE + Amort.
11.14813.52415.918.276Feb 19, 2026

Evaluation results

MethodLinks
GrMoE + Amort.Scale=2.7B, N=322026.0211.50.09201.38
Soft MoEScale=2.7B, N=322026.0211.80.148201.48
Softmax Top-2Scale=2.7B, N=322026.0212.10.342651.04
GrMoE + Amort.Scale=1.3B, N=162026.0213.80.08101.42
Soft MoEScale=1.3B, N=162026.02140.128151.51
Softmax Top-2Scale=1.3B, N=162026.0214.20.298551.08
Expert ChoiceScale=1.3B, N=162026.0214.40.07201.85
SwitchScale=1.3B, N=162026.0214.60.185300.82
GrMoE + Amort.Scale=350M, N=82026.0218.10.07401.29
GrMoEScale=350M, N=82026.0218.30.08901.34
Soft MoEScale=350M, N=82026.0218.40.112101.42
vMF-GateScale=350M, N=82026.0218.50.198201.18
Softmax Top-2Scale=350M, N=82026.0218.70.241401.12
Expert ChoiceScale=350M, N=82026.0218.90.06801.71
SwitchScale=350M, N=82026.0219.10.152200.89
HashScale=350M, N=82026.0220.30.03802.05