ModelsMedium importance

SenseTime launches SenseNova-Loop-9B, a model that runs the same block of layers twice

SenseNova-Loop-9B-Preview, released on 10/30 under the Apache 2.0 license, is multimodal (text and image input) and was trained with a context of up to 262 thousand tokens. It has about 9 billion parameters, but runs layers 4 to 27 twice with the same weights: 56 layer executions from 32 stored layers, the equivalent of about 14.6 billion parameters at inference. There is already support via pull requests in vLLM and SGLang.