SSyncDev

SenseNova-Loop-9B-Preview

SyncDev model directory · Profile updated Nov 18, 2026

Compare with another model

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.

Key specifications

Specification score
40.1
Context window
262.14K
License
Apache-2.0

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

Pricing conditions: No API pricing stated; open weights for self-hosting

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

Function callingVisionOpen weights

What is SenseNova-Loop-9B-Preview?

Can a 9B model think like a deeper one without carrying the extra weights? SenseNova-Loop-9B-Preview tests that idea in public. A contiguous block of its decoder runs twice with the same weights, so each token passes through 56 layer executions while the checkpoint stores only 32 distinct layers.

SenseTime released the preview at the end of October 2026 for long-context and agentic workflows: document understanding, coding and tool use. It reads text and images and writes text.

The architecture in brief

  • Shared looped block. 24 decoder layers execute twice in sequence.
  • Two-stream mHC. Manifold-constrained hyper-connections keep two residual streams in the looped section and learn how information moves between them.
  • Hybrid attention. Linear and full attention alternate in a repeating four-layer pattern, which keeps long prompts affordable.
  • Long training sequences. Recipes mix text, images, code and tool trajectories at up to 262,144 tokens.
Spec Value
Hub repo sensenova/SenseNova-Loop-9B-Preview
Stored parameters about 9B
Effective inference parameters about 14.6B
Input / output Text and images / text
Max training length 262,144 tokens
Licence Apache-2.0

The trade-off nobody mentions in the headline

Looping saves memory, not compute. The weights are 9B, but the forward pass behaves more like a 14.6B model, so expect latency and FLOPs closer to the larger number. If your bottleneck is VRAM on a single card, that is a good trade. If it is tokens per second at scale, it may not be.

Dates and the final checkpoint

The Hub repository is dated 28 October and the release note says 30 October. A non-preview SenseNova-Loop-9B repository appeared on 17 November behind gated access; this page covers the preview, whose specification is fully documented. Check the final repo's terms and changes before switching.

Testing it against your work

Pick long, messy inputs: a 200-page PDF, a repository with a failing test, a multi-step tool task. Measure accuracy, wall-clock latency and peak memory against a conventional model of similar size, such as Qwen3.5 9B. Use SenseTime's documented deployment settings, since looped models are sensitive to serving configuration. More options sit in the image-text-to-text directory.

How to evaluate SenseNova-Loop-9B-Preview

Compare the model on the work you need to complete.

Tasks and inputs

SenseNova-Loop-9B-Preview is listed for Image-text-to-text, Text 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: 262.14K 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

A block of 24 decoder layers is executed twice with shared weights, giving 56 layer executions from 32 stored layers.

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

sensetime

SenseNova-Loop-9B

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.

524.29K context

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

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