SenseTime has released SenseNova-Loop-9B, an open model with about 9 billion parameters that runs a block of its layers twice to think more deeply without growing larger. The weights and code went live on 17 November 2026 under the Apache 2.0 license, with downloads on Hugging Face gated behind an access request.
Quick Take
- Size: about 9B parameters, 32 physical decoder layers expanded to 56 layer computations.
- Inputs: text and images in, text out, with a context window of up to 512K tokens.
- Memory: the official 4-bit weights take about 4.9 GB, and SenseTime says deployment can need under 10 GB of GPU memory.
- Benchmarks (self-reported): 41.36 on NL2Repo versus 44.60 for Qwen3.8-27B, and 65.31 on AIDABench versus 65.01.
- License: Apache 2.0, gated download.
In plain English
Most AI models get smarter by getting bigger, which makes them expensive to run. SenseTime tried another trick: reuse the same middle layers twice, so the model does more thinking per word while keeping its size small. The result is a model that fits on a consumer graphics card and, by SenseTime’s own numbers, comes close to a model three times its size on some agent and coding tests.
What happened
SenseTime published the release on its website and GitHub. It follows a preview released on 31 October. According to the model card, as summarized by ORIENS, the final version scores higher than the preview on all seven benchmarks shown, with GAIA rising from 37.86 to 69.90. All figures come from SenseTime and have not been verified by a third party.
Background
Small open models that run locally are in high demand, and Alibaba’s Qwen3.8-27B has become a common yardstick, as seen in projects like local Qwen 3.8-27B builds. SenseTime’s design also uses manifold-constrained hyper-connections, the idea DeepSeek described in its mHC research.
How it works
- Shared loop: layers 4 to 27 form a weight-sharing block that runs twice in each forward pass.
- Dual-stream mHC: learned mappings route and fuse information across two residual streams inside the loop.
- Hybrid attention: layers come in groups of four, three with linear attention and one with full attention.
- Training: the recipe covers text, images, code and tool-use trajectories.
- Deployment: with 4-bit weights and an 8-bit KV cache, inference at a 256K context needs about 15 GiB of GPU memory. An official AWQ 4-bit version is available from day one, with support for Transformers, vLLM and an SGLang pull request.
Why it matters
For developers, a capable multimodal model with a long context that runs on one consumer GPU makes local agents and on-premises deployments cheaper.
For companies, Apache 2.0 allows commercial use, though the gated download adds a step.
For researchers, looped layers are a test of whether depth can partly replace parameter count.
What’s still open
- Independent benchmark results.
- Real-world speed, since running layers twice adds compute per token.
- How the model compares with the newest small open models outside the Qwen family.
Sources
- SenseTime, “SenseNova Loop-9B Official Release: Deeper Computation with Fewer Parameters,” 17 November 2026
- OpenSenseNova/SenseNova-Loop repository on GitHub
- Hugging Face model page, sensenova/SenseNova-Loop-9B
- ORIENS, report on the SenseNova-Loop-9B release, 17 November 2026
i10x Perspective
Looping layers is a clever answer to a practical problem: memory, not compute, is what limits most local deployments. If independent tests confirm SenseTime’s numbers, expect others to copy the idea. More in our Models section.



