Skip to main content

SenseNova-Loop-9B: A 9B Open-Weight Model That Runs Its Layers Twice

SenseNova-Loop-9B is a 9-billion-parameter text and image model released on Hugging Face under Apache 2.0 on 17 November 2026. It runs 24 of its 32 decoder layers twice per forward pass, giving about 14.6B inference-equivalent parameters. The model card says a 4-bit build is about 4.9 GB and needs roughly 15 GiB of GPU memory at a 256K context.

On this page

What SenseNova released on 17 November

The SenseNova organisation on Hugging Face has published SenseNova-Loop-9B, a model that takes text and images as input and produces text. The model card dates the full release to 17 November 2026, following a preview version on 31 October, and says both the weights and the code are available to the community.

The project is licensed under Apache 2.0, and the companion GitHub repository carries the same license. The Hugging Face repository is gated, which means you have to agree to share your account details before downloading. There is no paper and no separate launch post: the card and the README are the documentation.

How a looped decoder gets more from 9B parameters

The unusual part is the architecture. The model stores 32 distinct decoder layers, about 9B parameters in total. On each forward pass, layers 0 to 3 run once, then the block of layers 4 to 27 runs twice, then layers 28 to 31 run once. That makes 56 layer executions from 32 sets of stored weights, which the card describes as about 14.6B inference-equivalent parameters.

Put simply, the model spends extra compute time instead of extra storage. The download and the memory needed for weights stay at 9B scale, while each token passes through more layers than a normal 9B model would allow. Two loops is the default. The card reports that pushing to three or four loops lifts some scores further, for example a gain of 8.10% on LiveCodeBench V6 at four loops compared with one.

The trade-off is speed: running 24 layers twice means more work for every token, so a looped 9B model will not be as fast as a plain one of the same size. The card also lists a context length of 512K tokens and a hybrid attention design that mixes three linear-attention layers with one full-attention layer in repeating groups.

The benchmark claims against Qwen3.8-27B

SenseNova compares the model with Qwen3.8-27B, which has three times as many stored parameters. According to the card, Loop-9B ties it on AutomationBench at 36.67 and edges ahead on AIDABench, 65.31 against 65.01, while trailing on NL2Repo, 41.36 against 44.60.

The release is also a large step up from the preview. The card reports AutomationBench rising from 14.33 to 36.67, Terminal-Bench 2.1 from 40.45 to 59.55, and GAIA from 37.86 to 69.90.

All of these figures are self-reported, and no independent results exist yet. The card is candid about weak spots: the model may miss requirements when given many complex tool definitions or several constraints at once, its performance on coding and general agent tasks can be inconsistent, and complex web generation may need two or three rounds of refinement.

What it takes to run it yourself

This is where the release matters for local AI. The card says an official AWQ 4-bit version is available on launch day, with weights of about 4.9 GB. Its low-memory table estimates a minimum of roughly 10 GiB of GPU memory at a 64K context, 12 GiB at 128K, and 15 GiB at 256K, using 4-bit weights and an 8-bit key-value cache. That puts a multimodal model with a very long context within reach of a single consumer graphics card.

Software support is still catching up. Transformers works today with the trust-remote-code flag and a pinned library version. Support in vLLM and SGLang depends on pull requests that the card links but does not describe as merged. The card itself adds that its deployment commands were adapted from the preview and will be updated. It gives no instructions for llama.cpp, so phone and laptop users who rely on GGUF tooling should expect to wait for community support of the looped design.

If the scores hold up in outside tests, looping is a promising route for on-device AI: more capability from the same download size, paid for in inference time. SenseNova says it plans to apply the approach to larger models next.

CuriousLM runs supported AI models locally on your device. Try CuriousLM.