How much does SenseNova-Loop-9B-Preview cost?

Free under Apache 2.0. No hosted API listed for this model.

What are the downsides of SenseNova-Loop-9B-Preview?

No technical report; no vision, knowledge or reasoning benchmarks; one of eight benchmarks is SenseTime's internal test; 38.1 GB checkpoint stored mostly in FP32; examples use four GPUs; vLLM and SGLang support only in unmerged pull requests.

How much does SenseNova-Loop-9B-Preview cost for a team or in production?

The 38.1 GB checkpoint and four-way tensor parallelism in every serving example put it on multi-GPU servers, not single cards. Each token costs 56 layer passes, so per-token compute is closer to a 14.6B model than a 9B one.

Is SenseNova-Loop-9B-Preview safe for company data?

Apache 2.0 with no use restrictions; self-hosted only. Loading requires remote code, which runs SenseTime's Python on your machine. No training-data statement beyond long-sequence multimodal and tool-trajectory recipes.

What are the best SenseNova-Loop-9B-Preview alternatives?

The closest alternatives we have reviewed are ZDTaichu5.0-9B, Qwen3.8-27B, MiniCPM-V 4.7 (35B-A3B and 1B) and Nanbeige4.3-33B.

SenseNova-Loop-9B-Preview review: pricing and verdict (2026)

What is SenseNova-Loop-9B-Preview?

SenseNova-Loop-9B-Preview is an open multimodal model from SenseTime's SenseNova team, released on 30 October 2026 under Apache 2.0. It reads text and images and writes text, and is pitched at long-context agent work: documents, code, spreadsheets, slides and tool calls.

Its distinctive feature is a looped decoder. The model stores 32 decoder layers, but layers 4 to 27 run twice in sequence with the same weights, so each token passes through 56 layer computations. SenseTime describes this as about 9 billion stored parameters doing the work of about 14.6 billion. Inside the loop, two parallel residual streams are mixed by learned, constrained mappings (manifold-constrained hyper-connections, or mHC), and attention alternates three linear-attention layers with one full-attention layer.

It was trained on sequences of up to 262,144 tokens and supports multi-token prediction for faster decoding. The technical report is listed as "TBD", and the Transformers implementation's publication details are still to be announced.

Is SenseNova-Loop-9B-Preview worth it?

SenseNova-Loop-9B-Preview's launch chart is strong for a 9B model and narrow in what it tests. Across eight benchmarks it leads Qwen3.5-9B on every one, often widely: 66.89 against 41.63 on WorkBuddy Bench, 58.87 against 40.78 on AIDABench and 13.72 against 1.05 on AutomationBench. It also tops the larger Qwen3.5-27B on six of the eight. All eight are agent, office or database tasks; one, RaccoonPPTBench, where it scores 70.70, is SenseTime's internal slide-generation test. There are no vision, knowledge or reasoning benchmarks, despite the model reading images, and no independent results yet.

The chart has soft spots. On Terminal Bench 2.1 its 40.45 merely ties Qwen3.5-35B-A3B and Qwen3.6-35B-A3B, and trails Qwen3.5-27B at 47.19 and the 3B Nanbeige4.2 at 49.44. On τ³-Banking Qwen3.5-27B leads, 19.59 to 18.56.

The looped design is the reason to watch it. Running layers 4 to 27 twice gives the depth of a larger model while storing only 32 layers, but it does not save compute: each token still costs 56 layer passes, roughly a 14.6B model's work by SenseTime's own description. The saving is in memory for weights, not in speed.

In practice even that saving is not yet visible. The Hugging Face repository holds 38.1 GB across nine shards, and its metadata lists 9.41 billion parameters stored as FP32, although the card describes a BF16 backbone. Every serving example uses four GPUs with tensor parallelism, vLLM and SGLang support sit in unmerged pull requests, and a standard runtime that runs each layer once would compute a different network, so community quantisations need care. The preview label is earned.

Where it earns its place

An ML researcher at a university lab working on efficient architectures
Study how weight-shared loops trade parameters for depth
Load the checkpoint in Transformers 5.18.0 with trust_remote_code and inspect the loop settings, which define layers 4 to 27 as the repeated block with 2 passes. SenseTime's torchtitan fork includes SFT, DPO and MTP training recipes for single-node 8-GPU runs.
An AI platform team at a consultancy automating slide decks and spreadsheet analysis
Trial a small open model for office-agent tasks before committing to a larger one
Serve it with LightLLM on four GPUs as the card shows, then compare it with Qwen3.5-27B on your own tasks. Its 66.89 WorkBuddy and 70.70 RaccoonPPTBench scores are SenseTime's, and the second is an internal test.
A self-hosting engineer at a 50-person software firm
Run a local multimodal agent model on one workstation GPU
Wait for now: the official weights are 38.1 GB, the card's examples use tensor parallelism across four GPUs, and no quantised build from SenseTime exists. Any community quantisation must keep the loop and routing operators intact to compute the same model.

Before you commit

Setup: Install transformers 5.18.0 and PyTorch 2.11.0 with CUDA 13.0, then load with AutoModelForCausalLM and trust_remote_code; the model code ships in the repository. For serving, build vLLM from pull request 64832 or SGLang from pull request 47281, or use the LightLLM Docker image.
Integrations: OpenAI-compatible chat endpoints from LightLLM, vLLM or SGLang; a thinking mode switched with enable_thinking; MTP speculative decoding now, with DSpark and DFlash2 draft models still in training. Weights are mirrored on ModelScope.
Data and privacy: Apache 2.0 with no use restrictions; self-hosted only. Loading requires remote code, which runs SenseTime's Python on your machine. No training-data statement beyond long-sequence multimodal and tool-trajectory recipes.
Cost at scale: The 38.1 GB checkpoint and four-way tensor parallelism in every serving example put it on multi-GPU servers, not single cards. Each token costs 56 layer passes, so per-token compute is closer to a 14.6B model than a 9B one.
Limits: Preview release with no technical report. Config allows 262,144 positions, the training length; English and Chinese are listed. Runtimes that execute each layer once compute a different network, so quantised or ported builds must reproduce the loop.

What’s good

Leads Qwen3.5-9B on all eight launch benchmarks (self-reported)
66.89 on WorkBuddy Bench and 58.87 on AIDABench, above Qwen3.5-27B
Looped design gives 56 layer passes from 32 stored layers
Apache 2.0, with fine-tuning and draft-model training scripts

What’s not

No technical report; no vision, knowledge or reasoning benchmarks
One of eight benchmarks is SenseTime's internal test
38.1 GB checkpoint stored mostly in FP32; examples use four GPUs
vLLM and SGLang support only in unmerged pull requests

Best for

Research teams studying looped and weight-shared architectures, or evaluating small open models for office-agent work, who can tolerate preview-grade tooling.

SenseNova-Loop-9B-Preview alternatives

ZDTaichu5.0-9B

Zidong Taichu (CASIA and Wuhan AI Research)Vision & multimodalOpen weights

3.8/5

A 9B vision-language model from China's Zidong Taichu team that trades a little OCR for spatial reasoning and tool use.

Qwen3.8-27B

Alibaba (Qwen)General purposeOpen weights

4.5/5

Alibaba's Apache 2.0 dense model: 27 billion parameters, image and video input, and coding scores that rival models many times its size.

MiniCPM-V 4.7 (35B-A3B and 1B)

OpenBMBVision & multimodalOpen weights

4.0/5

OpenBMB's Apache 2.0 vision models: a 35B mixture-of-experts that leads on documents and a 1B for phones.

Nanbeige4.3-33B

BOSS Zhipin (Nanbeige LLM Lab)CodingOpen weights