Huginn-3.5B

"Scaling up Test-Time Compute with Latent Reasoning: A Recurrent Depth Approach" (NeurIPS 2025): a 3.5B depth-recurrent language model pretrained from scratch on 800B tokens on ORNL Frontier that scales reasoning by unrolling a recurrent block at inference — latent test-time compute with no chain-of-thought tokens. Joint work by Jonas Geiping (ELLIS Tübingen / MPI-IS / Tübingen AI Center) with Tom Goldstein's UMD group; weights are hosted under UMD's HF org.