SenseNova-Loop-9B
SenseTime released the final SenseNova-Loop-9B, a roughly 9B looped multimodal model with a 512K context, as open weights and code, following the October preview.
The final checkpoint replaces the October preview with updated weights and a longer 512K-token context. It keeps 32 decoder layers and runs the shared middle block (layers 4 to 27) twice, giving 56 layer computations per forward pass, with two residual streams using mHC. In SenseTime's own tests GAIA rises from 37.86 for the preview to 69.90, and Terminal-Bench 2.1 from 40.45 to 59.55. Cambricon reported launch-day support on its MLU platform.
- Date
- Tuesday 17 November 2026
- Lab
- SenseTime
- Kind
- open-weights
- Access
- open weights
Figures
| Measure | Value | Measured by |
|---|---|---|
| GAIA | 69.90 37.86 for the Preview | company |
| Terminal-Bench 2.1 | 59.55 40.45 for the Preview | company |
| AIDABench | 65.31 Qwen3.8-27B scores 65.01 | company |
| NL2Repo | 41.36 Qwen3.8-27B scores 44.60 | company |
| Context window | 512K tokens | company |
| 4-bit AWQ weights size | about 4.9 GB about 15 GiB GPU memory for 256K context with 8-bit KV cache, per model documentation | company |
Benchmarks are developer-reported and not independently reproduced. Pandaily says the project is under Apache 2.0, and deployment docs still carry Preview-derived configurations.
Sources
- pandaily.com/sensetime-sensenova-loop-9b-final-cambricon-mlu-support
- technode.com/2026/11/18/sensetime-releases-open-weight-loop-9b-model-that-reuses-layers-fo
This record was checked against its sources on 18 November 2026. How we check