SenseNova-Loop-9B-Preview

A 9B vision-language model that runs 24 of its 32 layers twice. Apache 2.0, 256K context, but it needs custom code.

In short. SenseNova-Loop-9B-Preview reads text and images and is aimed at long documents, coding and tool use. Its trick is to reuse part of its own layers twice, so it thinks deeper without storing more parameters. It is an early preview and needs special code, so popular local tools cannot run it yet.

Summary

SenseNova-Loop-9B-Preview is a multimodal model from SenseTime's SenseNova team, released on 30 October 2026 under the Apache 2.0 license. It accepts text and images and writes text, and SenseTime aims it at long-context and agentic work such as document understanding, coding and tool use. It is labeled a preview, and the technical report and the notes on the Transformers implementation are still marked "TBD" or "to be announced".

Its defining feature is a looped decoder. It stores 32 distinct decoder layers, but layers 4 to 27 run twice with shared weights, so each token passes through 56 layer executions. SenseTime counts this as about 14.6 billion "inference parameters" against about 9 billion stored ones. The loop sits inside two residual streams whose mixing is learned (two-stream manifold-constrained hyper-connections). Attention alternates three linear-attention layers with one full-attention layer. The hidden size is 4,096, there are 16 attention heads with 4 key-value heads, the vocabulary has 248,320 tokens and the training context is 262,144 tokens. The repository includes weights for multi-token prediction to speed up decoding, and two draft models, DSpark and DFlash2, are listed as still in training.

SenseTime shows its results only as a chart, which compares the model with Nanbeige4.2-3B, IQuest-Coder-V1-40B-Loop-Thinking and Qwen3.5 and Qwen3.6 models. It leads there on BIRD-Dev, a text-to-SQL test (71.38 against 69.84 for Qwen3.5-27B), and on WorkBuddy Bench (66.89 against 61.98 for the best Qwen model). On Terminal-Bench 2.1 it scores 40.45, below Qwen3.5-27B (47.19) and Nanbeige4.2-3B (49.44). These are SenseTime's numbers. The hosted files are stored mostly in FP32 (38.1 GB including the multi-token-prediction file) although the card describes a BF16 backbone, and the card's example loads them as BF16. The model needs custom code: the card lists Transformers with remote code, a LightLLM recipe for four GPUs, and open pull requests for vLLM and SGLang, and it does not list llama.cpp, Ollama or LM Studio.

Good for

  • Experiments with looped, multimodal agent models on one GPU with about 24 GB of memory
  • Research on looped transformers and hyper-connections, since training scripts and draft-model training code are public
  • Prototypes that read documents and charts, call tools and write code in one small model
  • Long inputs of up to 256K tokens, within the memory you have for the cache

Watch out for

  • It is a preview with no technical report yet, and it runs only through custom code or unmerged pull requests
  • Every token runs 56 layer executions, so it is slower than a standard 9B model, and the card says nothing about speed
  • There are no quantized builds, so by our estimate the BF16 weights need about 22 GB of memory
  • The benchmark chart is SenseTime's own, and in it the model trails Nanbeige4.2-3B on Terminal-Bench 2.1
  • The card says BF16 while the hosted files are mostly FP32, so the download is about 38 GB

Performance claims come from the developer and have not been independently verified.

How much memory does it need?

Estimates for the SenseNova-Loop-9B-Preview variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.

Variants Parameters BF16 FP8 / INT8 4-bit (Q4) Fits on
SenseNova-Loop-9B-Preview FP32 9.66B 22 GB 11 GB 22 GB Gaming GPU

Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software. The developer states a minimum of 22 GB of GPU memory.

VRAM calculator

The license in plain words

One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.

More about Apache 2.0

See also

Frequently asked questions

What is SenseNova-Loop-9B-Preview?
SenseNova-Loop-9B-Preview reads text and images and is aimed at long documents, coding and tool use. Its trick is to reuse part of its own layers twice, so it thinks deeper without storing more parameters. It is an early preview and needs special code, so popular local tools cannot run it yet.
Can I use SenseNova-Loop-9B-Preview commercially?
Yes. SenseNova-Loop-9B-Preview is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
Is SenseNova-Loop-9B-Preview open source?
SenseNova-Loop-9B-Preview is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
Can I run SenseNova-Loop-9B-Preview on my own computer?
Yes, on a good graphics card. The smallest variant, SenseNova-Loop-9B-Preview, needs about 22 GB of memory at 4-bit precision, which suits a 24 to 32 GB gaming GPU or a Mac with 32 GB or more.
When was SenseNova-Loop-9B-Preview released?
SenseNova-Loop-9B-Preview was released by SenseTime on October 30, 2026.
How long a context window does SenseNova-Loop-9B-Preview have?
SenseNova-Loop-9B-Preview supports a context window of up to 256K tokens.

Official links

Sources

Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.

Release date and facts last verified November 11, 2026.

Related models

Model Released Parameters Context License Smallest variant needs
Nanbeige4.2-3B Nanbeige LLM Lab 4.17B 256K Apache 2.0 (Permissive) 2.88 GB Laptop or phone
MiniCPM-V 4.7 OpenBMB 1.55B – 35.2B 3B active 256K Apache 2.0 (Permissive) 1.07 GB Laptop or phone
Qwen3.8-27B Alibaba Qwen 27.8B 256K Apache 2.0 (Permissive) 19 GB Gaming GPU
Gemma 4 Google DeepMind 5.1B – 30.7B 256K Apache 2.0 (Permissive) 3.52 GB Laptop or phone