NewsPublished:
SenseNova-Loop-9B-Preview Runs Its Own Layers Twice, and It Is Open Source
SenseTime's new Apache 2.0 multimodal model reuses a 24-layer block with shared weights to get 56 layer passes from 32 layers, aimed at agent work.
Most language models read your prompt through each of their layers exactly once. SenseTime's new open model takes a different route: it runs a big chunk of itself twice, back to back, with the same weights. The result is SenseNova-Loop-9B-Preview, released on October 30 under the Apache 2.0 license, and it's one of the more interesting small-model experiments I've seen this month.
How the loop works
According to the official model card, the model has 32 distinct decoder layers. Layers 4 through 27, a block of 24, are applied two times in a row, so every forward pass does 56 layer computations while only storing 32 layers' worth of weights. That's why SenseTime lists it at about 9B total parameters but roughly 14.6B "inference parameters": you get the depth of a bigger model without the memory bill for extra weights.
Two other pieces support the loop:
- Hybrid attention: in every group of four layers, three use linear attention and one uses full attention, which helps with long inputs
- Two-stream mHC (manifold-constrained hyper-connections): learned mappings that manage how information flows between two residual streams inside the looped section
What it's built for
SenseNova-Loop takes text and images in and produces text. SenseTime pitches it for long-context and agentic work: reading documents, writing code and calling tools over several steps. Training recipes go up to 262,144 tokens of context, and the model supports MTP for faster inference, with two more speed-up methods still in training.
Worth a look?
If you follow open models, the idea here, scaling compute depth instead of parameter count, is the real story. It's a preview, the technical report is still marked "TBD", and the Transformers implementation details are listed as coming later, so expect some rough edges. But for anyone tinkering with local models for creative or coding tasks, a 9B model that "thinks deeper" by reusing its own layers is a neat trick to try.
Source: SenseNova-Loop-9B-Preview model card on Hugging Face
Source: huggingface.co