SenseNova-Loop-9B-Preview
SenseTime open-sourced SenseNova-Loop-9B-Preview, a 9B looped Transformer that reuses a 24-layer block twice and that the company says matches models of about 27B parameters.
The model stores 32 decoder layers but runs 56 layer executions per forward pass, because layers 4 through 27 run twice with shared weights. It adds two-stream manifold-constrained hyper-connections (mHC) inside the looped section. It takes text and images and outputs text, with training sequences up to 262,144 tokens.
- Date
- Friday 30 October 2026
- Lab
- SenseTime
- Kind
- open-weights
- Access
- open weights
Figures
| Measure | Value | Measured by |
|---|---|---|
| Total parameters | about 9B about 14.6B inference parameters | company |
| mHC ablation gain on GPQA Diamond | +4.42 points ablation inside the company's own tests | company |
| mHC ablation gain on TeamBench | +13.34 points ablation inside the company's own tests | company |
| Layer executions per forward pass | 56 from 32 stored layers | company |
Weights are on Hugging Face under Apache 2.0. Benchmark results are the company's own and the technical report is still to come.
Sources
This record was checked against its sources on 1 November 2026. How we check