Is SenseNova-Loop-9B-Preview state of the art?Pending benchmarks.Announced Oct 30, 2026; no independent results yet.
SenseTime's open-weights looped multimodal model for long-context and agentic work. Of its 32 decoder layers, layers 4–27 run twice with shared weights (56 layer passes per token), and two-stream Manifold-Constrained Hyper-Connections (mHC) route information between passes. Takes text and images, returns text. Weights on Hugging Face under Apache-2.0; a preview, with the technical report still to come.
The details so far
- Lab
- SenseTime
- Announced
- Oct 30, 2026
- API
- Not yet
- Open weights
- Hugging Face (sensenova/SenseNova-Loop-9B-Preview, first published as SenseNova-6.8-Loop-9B-Preview)
- Parameters
- About 9.66B stored (9,656,250,480); SenseTime cites ~9B total and ~14.6B inference parameters with the loop
- Type
- Looped (parameter-shared) transformer, image-text-to-text; custom code (trust_remote_code)
- Loop
- 32 distinct decoder layers; layers 4–27 (24 layers) run twice with shared weights, 56 layer executions per forward pass
- mHC
- Two-stream Manifold-Constrained Hyper-Connections with Sinkhorn normalization (20 iterations)
- Attention
- Hybrid: 3 linear-attention layers then 1 full-attention layer in each group of 4
- Dimensions
- Hidden 4,096; FFN 12,288; 16 attention heads, 4 KV heads; 248,320-token vocabulary; Qwen3.5-style text config
- Context
- Trained at up to 262,144 tokens
- Checkpoint
- 38.1 GB; safetensors list 9.41B F32 and 0.24B BF16 parameters, while the card describes a BF16 backbone with FP32 mHC routing
- License
- Apache-2.0
- Serving
- LightLLM; SGLang support in an open pull request; fine-tuning via OpenSenseNova/torchtitan
- Status
- Preview; technical report TBD; MTP heads available, Dspark and Dflash2 still training
Tags marked unsure are not stated on the lab's own pages.
What would let us call it
Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.
- Publicly announced with a source
- Usable by the public (API or weights)
- Listed on an independent board we track
- Enough results for a rank in a category
- Holds across two snapshots (crowns only)
What the lab claims
Self-reported. Shown for context; not used in any verdict.
Lab claim: 71.38 on BIRD-Dev, 66.89 on WorkBuddy Bench, 58.87 on AIDABench, 48.89 on TeamBench, 40.45 on Terminal Bench 2.1, 18.56 on τ³-Banking, 13.72 on AutomationBench and 70.70 on RaccoonPPTBench (SenseTime's internal benchmark), from the model card chart.
Lab claim: ahead of Qwen3.5-27B on WorkBuddy Bench, RaccoonPPT Bench and AutomationBench.
Lab claim: mHC adds 4.42 points on GPQA Diamond and 13.34 on TeamBench.
Sources
REPOsensenova/SenseNova-Loop-9B-Preview on Hugging Facehuggingface.co/sensenova/SenseNova-Loop-9B-Preview · Oct 30, 2026REPOOpenSenseNova/SenseNova-Loop on GitHubgithub.com/OpenSenseNova/SenseNova-Loop · Oct 30, 2026LABSenseTime on Xx.com/SenseTime_AI/status/2116151078219858346 · Oct 30, 2026Tracking continues until an independent board lists it.