SmallThinker

IPADS's MoE family "natively trained for local deployment" — from-scratch pretraining co-designed with the on-device inference stack (fine-grained sparsity, pre-attention router for compute/IO overlap) rather than compressing a cloud model after the fact. Successor to the viral SmallThinker-3B-Preview (415 HF likes); weights now ship under the team's Tiiny-AI HF org.