SSyncDev

SenseNova-Loop-9B

SyncDev model directory · Profile updated Nov 19, 2026

Compare with another model

SenseNova-Loop-9B is SenseTime's 9B looped multimodal model: 32 layers run as 56, 512K context, ~15 GiB at 256K in 4-bit, open weights under Apache 2.0.

Key specifications

Specification score
58.5
Context window
524.29K
License
Apache-2.0

The specification score summarizes recorded features and limits. It is not a benchmark result.

Pricing conditions: Open weights under Apache 2.0; no hosted per-token price published.

Inputs and outputs

Recorded modality support shows what this model can accept and produce. An undocumented modality is shown as unknown rather than assumed to be unsupported.

Accepts

  • Text input
  • Image input

Produces

  • Text output

Capabilities

ReasoningFunction callingVisionOpen weights

What is SenseNova-Loop-9B?

Three weeks after the preview, SenseTime shipped the finished SenseNova-Loop-9B on 17 November 2026. The core idea is unchanged. Instead of stacking more layers, the model runs some of them twice. It has 32 distinct decoder layers, but layers 4 to 27 execute two times in a row, so every forward pass performs 56 layer computations while storing only 32 layers of weights. SenseTime puts the inference-equivalent size at about 14.6B parameters.

What changed is the scores, and by a lot.

Preview versus release

Benchmark Preview Release
AutomationBench 14.33 36.67
Terminal-Bench 2.1 40.45 59.55
GAIA 37.86 69.90
BIRD 71.38 72.34

SenseTime also compares it with Qwen3.8-27B, a model three times its stored size: a tie on AutomationBench (36.67), 65.31 against 65.01 on AIDABench, and 41.36 against 44.60 on NL2Repo. These are publisher numbers.

Architecture details

  • Looped block: layers [4, 28) run twice with shared weights; the first pass's output feeds the second.
  • Two-stream mHC: manifold-constrained hyper-connections carry two residual streams through the loop, with stream mixing constrained by 20 Sinkhorn normalisation iterations.
  • Hybrid attention: every group of four layers is three linear-attention layers plus one full-attention layer, and loop boundaries align with those groups.
  • Dimensions: hidden size 4,096, feed-forward 12,288, vocabulary 248,320, 16 full-attention heads with 4 KV heads.
  • Modalities: text and images in, text out.
  • Context: 512K tokens.

Any inference engine has to respect the loop range, pass count and mHC routing parameters, or outputs will be wrong. Use SenseTime's modelling code (modeling_sensenova_loop.py ships in the repository) rather than a generic loader.

Memory footprint

This is the practical headline. With 4-bit weights and an 8-bit KV cache, SenseTime says multimodal inference at 256K context needs about 15 GiB of GPU memory. An official AWQ 4-bit build was available on launch day, and the repository includes MTP weights. The Hugging Face repository is gated with automatic approval, so expect a click-through before download.

More loops, more quality

SenseTime's scaling study found that running the block three or four times improves results further (for example, up to 16.93% relative gain on an office benchmark at three loops) but defaults to two to keep inference fast. That is a dial worth knowing exists, though the release is tuned for two.

Testing it

The release targets agent, coding and data-analysis work, so test there. Run your text-to-SQL or repository tasks through the AWQ build on a single 16-24 GB GPU and compare against what you get today from a 27B-class model; the claim to verify is near-27B quality at a 9B memory bill. If you used the SenseNova-Loop-9B-Preview, rerun its failures first. For comparison, see Qwen3.5 27B, or browse image-text-to-text models.

How to evaluate SenseNova-Loop-9B

Compare the model on the work you need to complete.

Tasks and inputs

SenseNova-Loop-9B is listed for Image-text-to-text, Text generation, Coding & Code Generation. It accepts text, image and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.

Capacity and cost

Documented limits: 524.29K context. Measure a completed task using the prompts, documents and tools your application will use. Include reasoning and retries in the result.

Choosing a deployment

The weights are downloadable under Apache-2.0. Check runtime support and size the complete model and working context for your expected concurrency.

Frequently asked questions

Layers 4 to 27 run twice per forward pass with shared weights, giving 56 layer computations.

These records share a directory category. Check the source card, output type and deployment requirements before treating a model as a replacement.

Tencent

Hy4 Flash

Tencent Hy4 Flash is the low-cost Hy4 tier with open weights: a 220B MoE (~11.3B active) for agents and documents, with 1M context and image input.

1.05M context

shanghai-ai-lab

Sage S2.1 Flash

Sage S2.1 Flash is Shanghai AI Lab's open scientific multimodal agent model, reading molecules, crystals and proteins alongside text and images.

— context

Meituan

LongCat-2.5

LongCat-2.5 is Meituan's open-weight MIT-licensed multimodal MoE for long-horizon agents: screenshots, charts, video and multi-step coding.

— context

sensetime

SenseNova-Loop-9B-Preview

SenseNova-Loop-9B-Preview is SenseTime's looped 9B text-and-image model for long-context agentic work, running shared layers twice for extra depth.

262.14K context

Liquid AI

LFM3-VL-2B

LFM3-VL-2B is Liquid AI's open 2B vision-language model for images and video on-device, strong on documents, charts and GUI screenshots.

— context

NVIDIA

Nemotron 3.5 Super VL 120B A12B

Nemotron 3.5 Super VL 120B A12B is NVIDIA's open multimodal reasoning MoE: 12B active, 1M context, image and video input, switchable thinking.

1.05M context