Chinese AI company SenseTime announced on November 17, 2026 that it has officially released SenseNova-Loop-9B (Loop-9B below), an open-weight multimodal language model. The weights are on Hugging Face and the inference code is on GitHub, both under the Apache License 2.0. The model is the official version of SenseNova-Loop-9B-Preview, which came out on October 31.

24 of Its 32 Layers Run Twice With the Same Weights

What sets Loop-9B apart is that it reuses part of its layer stack. According to the model card, the model has 32 decoder layers, and layers 4 through 27 (24 layers) run twice with shared weights. The first four and last four layers run only once, so each forward pass executes 56 layers in total.

The model has about 9B total parameters, and SenseTime says inference is equivalent to about 14.6B parameters. In other words, the weights held in memory stay at the 9B class while the model computes more like a deeper network. It also uses mHC (manifold-constrained hyper-connections), an extension of residual connections.

The attention design is a hybrid that repeats a group of three linear-attention layers and one full-attention layer. The model takes text and images as input, outputs text, and supports a context length of up to 512K tokens.

Scores on Par With Qwen3.8-27B

SenseTime compares the model with Qwen3.8-27B, which has about three times as many parameters. The main figures in the model card are below.

BenchmarkLoop-9BQwen3.8-27B
AIDABench65.3165.01
AutomationBench36.6736.67
NL2Repo41.3644.60

Loop-9B comes out ahead on AIDABench, ties on AutomationBench, and trails on NL2Repo. It also improves sharply on the Preview: GAIA rose from 37.86 to 69.90, and Terminal-Bench 2.1 from 40.45 to 59.55. These are SenseTime's own measurements.

Two loops is the default, but the card also shows results with three and four loops. SenseTime says going from two to four loops adds only about 1 to 1.4 points.

4-Bit Weights Are About 4.9GB, but It Needs a Dedicated Setup

An official AWQ 4-bit quantized version is available from launch day, and the 4-bit weights are about 4.9GB. The KV cache can also be quantized to 8-bit or 4-bit. The model card gives the following GPU memory estimates.

Context length8-bit KV cache4-bit KV cache
64KAbout 10GiBAbout 9GiB
128KAbout 12GiBAbout 10GiB
256KAbout 15GiBAbout 12GiB

The software side is not fully ready yet. Running it with Transformers requires loading the model's own custom code, and vLLM and SGLang need patched branches that have not been merged upstream. SenseTime's announcement says that, as of this writing, the vLLM patch supports only unquantized text inference and does not support the low-memory AWQ 4-bit setup. Downloading from Hugging Face also requires agreeing to share user information.

SenseTime lists known limitations, including following complex instructions with many tool definitions and inconsistent results on coding and agent tasks.

A Notable Attempt to Get More Depth From a Small Model

We think reusing layers to deepen computation without adding memory could be an attractive option for developers who want to run LLMs on their own GPUs. If multimodal, long-context inference fits in roughly 9GiB to 15GiB, it should come within reach of consumer GPUs.

That said, because the layers run twice, inference may be slower than a conventional 9B model. SenseTime has not published speed figures, and quantized support in the major inference engines is still in progress. Even so, the Apache 2.0 license allows commercial use, and once inference engines catch up, the model is likely to become a strong candidate for local LLM use.

Summary

SenseTime has officially released SenseNova-Loop-9B, a 9B model that runs its middle 24 layers twice, as open weights under Apache 2.0. It gives developers and companies a new option for running multimodal models locally. The next things to watch are official support in vLLM and SGLang and how far quantized support progresses.