Eigen RadarAI
Analysis

SenseNova releases Loop-9B with reusable model layers

China’s SenseNova released the official Loop-9B model on November 17 after an October preview. The model processes text and images and returns text, reusing a shared block of layers to extend computation. New weights and code accompany the release. The public model page describes licensing, download conditions and limitations for complex instructions, coding and longer agent tasks.

Artificial Intelligence··Morning
Overhead view of an AI evaluation table with a smartphone clamped in a copy stand above a grid of printed test photographs and unlabelled chart sheets.

Official weights follow the preview

China’s SenseNova, the model project from AI company SenseTime, released the official Loop-9B weights and code on November 17. An October 31 preview had introduced the same looped architecture. The new release retains that structure and adds new weights for coding, data analysis and agent tasks. It accepts images alongside text and produces text responses.[1], [2]

The same layers run twice

Loop-9B stores about 9 billion distinct parameters, the values learned during training. Of its 32 separate decoder layers, a shared middle block of 24 runs twice, giving 56 layer computations in one forward pass. Both passes use the same weights; the first pass supplies the state processed by the second. Two parallel information streams are connected through learned mappings. The preview and official release both use two loops by default.[1]

Download conditions and task limits

The project carries an Apache 2.0 licence. Its public Hugging Face page requires users to log in and accept information-sharing conditions before downloading files. The model card reports difficulties with instructions containing many tool definitions or constraints, and inconsistent performance on long coding and agent tasks. Complex web-generation work can require additional rounds of feedback and refinement. The page lists no hosted inference provider for the model.[1]

References

  1. News sourceSenseNovaSenseNova Loop-9B releases a shared-layer multimodal model↩1↩2↩3
  2. News sourceIT之家SenseNova releases Loop-9B with shared model layers↩