SenseNova-Loop-9B

SenseNova-Loop-9B (released as SenseNova-Loop-9B-Preview) is an open-weight multimodal language model published by SenseTime's SenseNova team on October 30, 2026 under the Apache-2.0 license. Its defining idea is a looped decoder: instead of stacking ever more layers, a contiguous block of decoder layers is executed twice with shared weights, giving the model a deeper computation path without duplicating parameters.

The model accepts text and images and generates text, and it targets long-context and agentic workflows such as document understanding, coding, data analysis, slide/PPT design and tool use. SenseTime says that in its own tests the 9B model beats much larger dense models (for example Qwen 3.5 27B) on programming, office and agent tasks.

Key points

1. Looped block: layers 0–3 run once, layers 4–27 (24 layers) are applied twice in sequence with the same weights, then layers 28–31 and the output head finish. Result: 32 distinct layers, 56 layer executions per forward pass.
2. Two-stream mHC: manifold-constrained hyper-connections route and merge information between two residual streams inside the looped section, stabilizing the repeated computation.
3. Hybrid attention: linear attention and full attention are interleaved in a repeating four-layer pattern to keep long-context cost manageable.
4. Scale: ~9B stored parameters, ~14.6B "inference parameters" equivalent; BF16 backbone with FP32 routing parameters.
5. Long context: training recipes include text, images, code and tool trajectories at up to 262,144 tokens.
6. Acceleration: Multi-Token Prediction (MTP) is available now; Dspark and Dflash2 speculative decoders are still training.

Why it matters

Looped / recurrent-depth transformers are a way to trade compute for parameters: a small model that "thinks twice" per token. SenseNova-Loop is one of the first open, production-oriented releases of this idea from a major Chinese lab, making it useful for research on parameter-efficient reasoning and for teams that want a small multimodal agent model they can self-host.

How to use it

1. Download the weights from the Hugging Face model page; code lives in the OpenSenseNova/SenseNova-Loop GitHub repo.
2. Load it with Hugging Face Transformers using trust_remote_code (the looped architecture ships as a custom implementation), and follow the recommended deployment environment in the model card for best results.
3. Enable MTP for faster inference where supported.
4. Good fits: long-document Q&A, multimodal data analysis, coding assistants, PPT/motion design agents and tool-calling workflows on a single GPU.
5. Note it is a Preview: the technical report and official Transformers integration are still "TBD", so validate on your own tasks before production use.