Encoder backbone

Laya's English checkpoint is ModernBERT-large fully fine-tuned to 421M parameters; Julia 1 starts from mmBERT-small; opendecider-nano uses Ettin-encoder-400m. For all of them the encoder does the language work and the decision head reads options from its hidden states in one pass.

Different backbones pull different trade-offs. ModernBERT-large is strong at English, the mmBERT base covers 100+ languages at 322M parameters, and mmBERT-small keeps a 32,768-token context available at 144M — none of which implies zero-shot decision accuracy, which still comes from the head and its training.