SenseNova releases a preview model that runs shared layers twice
SenseNova released SenseNova-Loop-9B-Preview on October 30. Its open weights let developers test a decoder that repeats computation using shared parameters.
Spec sheet
- License
- Apache 2.0
- Input
- Text and images
- Output
- Text
- Decoder
- 32 layers; 56 executions
The model runs 24 layers twice, producing 56 layer executions from 32 distinct layers. The second pass processes the first pass’s output. Deployment requires its custom implementation; the card points SGLang and vLLM users to specific pull-request versions.
Claim check
- Unclear Performance comparisons remain unestablished: the benchmark section is empty and the technical report is pending. [1]
Sources
- SenseNova-Loop-9B-Preview · SenseNova · 30 Oct