# SenseTime opens SenseNova-Loop-9B, a model that runs layers twice

> The Apache 2.0 preview reuses 24 of its 32 layers twice per token, but the technical report and readable benchmarks are not out yet.

URL: https://claudegpt.si/news/sensenova-loop-9b-preview  
Published: 2026-11-01T02:50:32.695Z  
Author: ClaudeGPT Si Newsroom  
Section: AI Models  
Topics: SenseTime, Open-weight models, Hugging Face, Language models  
Publisher: ClaudeGPT Si

## Key takeaways

- SenseTime published SenseNova-Loop-9B-Preview on Hugging Face under Apache 2.0.
- Layers 4 to 27 of its 32 layers run twice, giving 56 layer passes, with about 9B parameters stored.
- The technical report is still marked TBD and benchmarks appear only as an image.

SenseTime has put SenseNova-Loop-9B-Preview on [Hugging Face](https://huggingface.co/sensenova/SenseNova-Loop-9B-Preview) under the Apache 2.0 license. The model reuses part of its own network twice for every token, which gives it more depth than its size suggests. [Pandaily reported](https://pandaily.com/sensetime-sensenova-loop-9b-preview-looped-transformer-mhc-open-source-apache-2) the release on November 1.

It is a multimodal model: it takes text and images and writes text. The model card says it is aimed at long documents, coding and tool use.

## How a looped model works

The model has 32 distinct decoder layers, the blocks that process each token. Layers 4 through 27 run twice with the same weights. That adds up to 56 layer passes from 32 layers' worth of parameters.

The card lists about 9 billion parameters in total and about 14.6 billion at inference, because the repeated layers do extra work. A looped design trades more compute per token for fewer weights to store.

## Other design choices

- Two-stream mHC: learned mappings control how information moves between two residual streams inside the looped section.
- Hybrid attention: linear and full attention layers alternate in a repeating four-layer pattern.
- Context: training covered sequences up to 262,144 tokens.
- Speed: multi-token prediction is available. Two further speed-up methods, Dspark and Dflash2, are still in training.

## Running it

The card says the model works with Hugging Face Transformers (tested with version 5.18.0 and PyTorch 2.11.0), LightLLM and SGLang. Transformers needs `trust_remote_code=True`, which runs code from the model repository, so read it first.

## What we don't know yet

The technical report is listed as "TBD", and the card shows benchmark results only as an image, with no numbers in text. Until SenseTime publishes the report or independent tests appear, there is no way to say how it compares with other 9B models. The "Preview" label also suggests the weights may change.

## Sources

- [SenseNova-Loop-9B-Preview model card](https://huggingface.co/sensenova/SenseNova-Loop-9B-Preview), Hugging Face
- [SenseTime open-sources SenseNova-Loop-9B-Preview](https://pandaily.com/sensetime-sensenova-loop-9b-preview-looped-transformer-mhc-open-source-apache-2), Pandaily