SenseTime's SenseNova-Loop-9B Runs Its Own Layers Twice, and It's Open Source
Primary source: sensenova/SenseNova-Loop-9B-Preview model card on Hugging Face (news entry dated 2026.10.30; checked Oct 30, 2026 ~7 PM ET). Code: OpenSenseNova/SenseNova-Loop on GitHub. Coverage: TechCity Authority (Oct 30, 2026).
Oct 30, 2026: Most AI labs get more reasoning out of a model by making it bigger. SenseTime's SenseNova team is trying something else: run part of the model twice. SenseNova-Loop-9B-Preview, released Oct 30 under Apache 2.0, stores about 9B parameters but does the compute of a deeper network. It's built for long documents, coding and agent work.
The core trick: a looped block
According to the model card, the model has 32 distinct decoder layers. Layers 0-3 run once. Then layers 4-27 (a 24-layer block) run twice in a row with the same weights, and the second pass takes the first pass's output. Finally, layers 28-31 and the output head finish.
The result, in the model card's words: "56 decoder-layer computations per forward pass while storing only 32 sets of decoder-layer weights." The card describes this as B(B(h)), the same block applied to its own output.
Key specs from the card:
| Property | Value |
|---|---|
| Total parameters | ~9B |
| "Inference parameters" | ~14.6B |
| Input / output | Text and images / text |
| Decoder layer executions | 56 per forward pass |
| Training context | Up to 262,144 tokens |
| Hidden / FFN dimension | 4,096 / 12,288 |
| Vocabulary | 248,320 |
| License | Apache 2.0 |
Two other design pieces
- Two-stream mHC: inside the looped section, "manifold-constrained hyper-connections" carry two residual streams, with learned mappings that decide how information is read, mixed and written back. The card says Sinkhorn normalization constrains the mixing, with 20 iterations in the documented config.
- Hybrid attention: each four-layer group has three linear-attention layers followed by one full-attention layer. The loop boundaries line up with these groups, so the pattern holds across both passes.
What it's for
SenseNova says the model is designed "for long-context and agentic workflows". That means tasks that mix document understanding, coding and tool use. A thinking mode can be turned on with enable_thinking=True.
The card includes a benchmark chart comparing the model with others on BIRD-Dev, WorkBuddy Bench, AIDABench, TeamBench, Terminal Bench 2.1, T³-Banking, AutomationBench and RaccoonPPTBench. These are self-reported results in an image. We're not repeating specific scores because we haven't independently verified them, and the technical report is listed as "TBD."
Running it
Per the model card:
- Tested setup: Transformers 5.18.0, PyTorch 2.11.0 with CUDA 13.0, BF16 weights
- Custom code: loading requires
trust_remote_code=True. That runs SenseNova's model code from the repo, so review it before using it in production. - Serving: documented for LightLLM (a four-GPU example) and SGLang through an open pull request (#47281), not a merged release
- Speedups: MTP (multi-token prediction) is "Available." DSpark and DFlash2 drafting are "Training in progress."
The card doesn't publish VRAM requirements.
Why it matters
"Looped" or weight-shared transformers have been a research idea for a while. A permissively licensed, multimodal, long-context release lets developers test whether extra depth without extra parameters actually pays off on agent tasks. It's still a Preview, with no tech report and no merged SGLang support, so treat it as something to experiment with, not a drop-in production model.
What's confirmed / not confirmed
Confirmed (model card / Hugging Face / GitHub):
- Release dated Oct 30, 2026 in the model card's News section. The Hugging Face repo was created Oct 28 and updated Oct 30 (Hugging Face API). The GitHub repo was created Oct 29 (UTC).
- Apache 2.0 license; text+image input, text output.
- 32 distinct layers, 24-layer block looped twice, 56 executions, ~9B total / ~14.6B inference parameters, up to 262,144-token training context.
- Requires trust_remote_code=True; SGLang support via PR #47281.
Not confirmed: - Benchmark claims. They're self-reported in a chart, and the technical report is "TBD." - How fast or slow the model is in practice compared with a standard 9B or 14B model. The card says speedups "depend on the workload, hardware, and serving configuration." - Official VRAM requirements. - That SenseTime announced on X. TechCity Authority reports this, but we didn't check the post.
Related
Up: AI
