SenseTime Releases an Open Model That Runs Part of Itself Twice

SenseTime has released the trained weights of SenseNova-Loop-9B-Preview, a model that takes in both text and images, under the Apache 2.0 license, which allows commercial use and modification. The repository went up on Hugging Face on Oct. 28, 2026, and the model card's own announcement line is dated Oct. 30, 2026.

The company describes the model as built for long documents and for software that works through tools, saying it is meant for jobs that combine reading documents, writing code and calling outside services.

The unusual part is how it is put together. The model card says 24 of the model's 32 decoder layers are run twice in a row, reusing the same weights on both passes, so a single pass through the model performs 56 layer computations rather than 32. SenseTime's stated purpose is a longer computation path without storing a second copy of the repeated layers. Inside that looped section, learned routing carries information along two parallel channels, a scheme the company calls two-stream manifold-constrained hyper-connections.

SenseTime published a performance chart alongside the release. Those comparisons are the company's own, and no independent evaluation of the model has been published. The card lists the technical report as still to be announced.

The company says its documented training recipes ran at sequence lengths of up to 262,144 tokens. The repository is tagged for English and Chinese.