SenseTime previews a multimodal model that reuses its decoder layers

SenseTime released a preview of a multimodal model that revisits a shared decoder block.

SenseTime released SenseNova-Loop-9B-Preview on October 30, publishing a model checkpoint and code for testing it. The model takes text and images and produces text. The release targets work such as inspecting documents, writing code and handling tasks that span several tool calls. Its architectural experiment is to spend more computation on an input by reusing part of the model rather than storing a new set of parameters for every layer that runs. [SenseTime repository](https://github.com/OpenSenseNova/SenseNova-Loop) · [Model card](https://huggingface.co/sensenova/SenseNova-Loop-9B-Preview)

The difference between stored weights and work at runtime

A conventional stack of transformer layers has a separate set of weights at each layer. SenseNova Loop passes through a shared block twice. The second pass uses the same weights on a different internal representation, so the model does additional work without doubling the weights for that block. A deployment team still has to pay for that second pass in inference time. Comparing this model with another nine-billion-parameter checkpoint by file size alone would miss that difference. [Architecture description](https://github.com/OpenSenseNova/SenseNova-Loop)

The trade is relevant when a team has limited accelerator memory but can tolerate a longer response. Loading the weights is one memory cost. Keeping intermediate attention state for a long document or several concurrent requests is another. A model small enough to load can still have very different throughput with short questions, high-resolution inputs or long prompt histories.

What a preview evaluation can establish

SenseTime's examples illustrate possible workflows and the model card provides the environment and prompt format. A practical document test would use pages with multi-column text, figures and small labels, then score answers against the actual page evidence. A coding test could use an unfamiliar repository and record whether the model found the relevant files and produced a patch that passed checks. Those tests measure different capabilities. A demo of one does not establish the other. [Model card and examples](https://huggingface.co/sensenova/SenseNova-Loop-9B-Preview)

The checkpoint is identified as a preview under Apache 2.0. It gives engineers an opportunity to compare memory, latency and task quality on hardware they control before committing to the architecture for a production workload. [SenseTime release](https://github.com/OpenSenseNova/SenseNova-Loop)