The official model card and repository confirm architecture, context, license, public weights and current serving paths.
SenseNova opens a looped multimodal model for long-context agents
2026-10-30Verified announcementTRIAL
SenseNova released ungated Apache-2.0 weights and code for SenseNova-Loop-9B-Preview, a text-and-image model for long-context and agentic workflows. It applies a 24-layer block twice, producing 56 layer executions from 32 distinct decoder layers, and was trained on sequences up to 256K tokens. Deployment currently relies on LightLLM main or an unmerged SGLang pull request.
Official · ReferenceSenseNova-Loop-9B-Preview model cardOfficial · ReferenceOpenSenseNova repository
Availability: Ungated BF16 weights and public code under Apache-2.0
What this means for you
For local-agent teams, shared-weight looping offers a testable compute-versus-parameter trade-off without a larger stored checkpoint. Preview status, custom model code and immature serving support make a bounded trial appropriate.
Evidence & related material
Technical details & scores
| Architecture | About 9B stored parameters, about 14.6B inference parameters; 32 distinct decoder layers and 56 executions; two-stream mHC |
|---|---|
| Modalities | Text and image input; text output |
| Context | Training sequences up to 262,144 tokens |
| Availability | Ungated BF16 weights and public code under Apache-2.0 |
| Serving | LightLLM main; SGLang requires an unmerged pull request |
| Time Basis | Official announcement gives 2026-10-30 without an hour; repository artifacts were public before the overlap start, so this is a late discovery. |
Technical impact 4
Engineering value 3
Adoption signal 2
Maturity 2
Verification cost 3
Risk 3