New model designs target efficient inference and scientific applications
Loopback Transformer reports lower validation loss with 40–75% of baseline parameters, and MMFlash claims lossless vision-language generation speedups up to 3.58×; LoopFill reports up to 5.85× decoding speedups for looped Transformers. UMM-JEPA reports improved visual benchmark results, while K2-Horizon emphasizes openness and training transparency. Bonbon reports molecular-interaction capabilities and prospective wet-lab hits, including a 61.5% hit rate; broad evaluation and independent replication are needed.
Sources (6)
Updated Nov 19, 2026