SenseTime releases SenseNova-Loop-9B-Preview, an open looped multimodal model

SenseTime's OpenSenseNova team released SenseNova-Loop-9B-Preview, an open-weight multimodal model built for long-context and agentic workflows, under the Apache 2.0 license with weights on Hugging Face. The model accepts text and images and generates text, aimed at use cases like document understanding, coding, and tool use.

Its core idea is a looped decoder: a block of 24 of the model's 32 decoder layers is run twice using the same shared weights, producing 56 total layer executions while storing only 32 layers' worth of parameters. The looped section also uses what SenseTime calls two-stream manifold-constrained hyper-connections, a mechanism that routes information across two residual streams using learned mappings constrained by Sinkhorn normalization. The decoder also interleaves linear-attention and full-attention layers in a repeating four-layer pattern, and was trained on documented recipes with sequences up to 262,144 tokens.

The model has roughly 9 billion total parameters but about 14.6 billion effective inference parameters due to the repeated computation. SenseTime published benchmark comparisons against other models on tests including BIRD-Dev, Terminal Bench 2.1, and RaccoonPPTBench, though it did not give the specific scores in the text accompanying those charts. The release includes deployment support for LightLLM, with community pull requests in progress for SGLang and vLLM, plus training recipes for fine-tuning and for speculative-decoding variants called DSpark and DFlash2, both still in training.

Why it matters

An open, loop-reuse architecture that boosts effective depth without adding stored parameters gives developers a new design to study for making smaller open models punch above their weight.

Source: SenseTime (OpenSenseNova on GitHub)