List your company

Language Modeling on Cosmopedia v2

6.49Final Perplexity (PPL)MoL K=5 top-3
6.4746.5826.696.798May 10, 2026

Evaluation results

MethodLinks
MoL K=5 top-3Params=104.0M2026.056.49-0.16
Sparse MoL 1+2of5Params=106.0M2026.056.52-0.13
Dense DeltaNetParams=103.4M2026.056.64-0.01
Dense softmaxParams=103.4M2026.056.65—
MoE-FFN 3of5Params=~106M2026.056.890.24