SSyncDev

SenseNova-Vision-7B-MoT

SyncDev model directory · Profile updated Nov 18, 2026

Compare with another model

SenseNova-Vision-7B-MoT is SenseTime's unified vision model that outputs boxes, masks, depth and normals by generation, with no task-specific heads.

Key specifications

Specification score
14.3
License
CC-BY-NC-4.0 (weights card); Apache-2.0 (code repo)

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
  • Image output

Capabilities

VisionOpen weights

What is SenseNova-Vision-7B-MoT?

A typical perception stack is a zoo: one detector, one segmenter, a depth network, a separate OCR engine, each with its own head and its own loss. SenseNova-Vision-7B-MoT folds that zoo into a single instruction-following model. You describe the task in words, and it generates the answer as text, as an image, or as both.

Three output shapes, one model

Output type Typical tasks What comes back
Structured text Detection, referring localisation, OCR, GUI grounding, keypoints, camera parameters Records with normalised coordinates
Dense image Depth, surface normals, point maps, binary or colour-coded masks Image-like maps
Mixed Multi-instance segmentation, grounded conversation segmentation Labels plus generated masks

Everything decodes back into standard boxes, points, masks and depth maps, so existing evaluation code still works. Instructions can also define task variants that no fixed benchmark schema covers, which is the more interesting property for product teams.

Licence: read this before you ship

The two official records disagree. The weights page marks the checkpoint CC-BY-NC-4.0, a non-commercial licence, while the code project states Apache 2.0. Assume the weights are non-commercial until SenseTime says otherwise in writing. The code being permissive does not change the weight terms.

What else arrived with it

SenseTime released the weights and inference code on 8 July 2026, along with the SenseNova-Vision-Corpus-50M dataset. A training pipeline for open datasets followed on 17 July, together with evaluation support for multi-view reconstruction and camera pose, and a data-preparation workflow update came on 22 July. The Hub repository itself is dated 29 June, a week before the announced release.

Getting it running

Setup uses the project's own scripts rather than a plain Transformers call: create the conda environment with the provided setup script, pull the checkpoint with huggingface_hub, then call the wrapper with a task name such as binary_seg plus a category and an image. A Gradio demo is included for quick inspection.

Where it fits

It suits research groups and internal tools that want one model for many dense tasks and can live with a non-commercial weight licence. If you need a production segmenter today, compare it with specialists in the image segmentation directory and with depth estimation models. Run your own images through every task you care about. Generalist models often look strong on average and weak on the one task you need.

How to evaluate SenseNova-Vision-7B-MoT

Compare the model on the work you need to complete.

Tasks and inputs

SenseNova-Vision-7B-MoT is listed for Depth estimation, Image segmentation, Object detection. It accepts text, image and produces text, image. Use the publisher’s input format and preprocessing for your chosen deployment.

Capacity and cost

Choose the output size and quality your design needs. Include drafts and revisions when budgeting an accepted image.

Choosing a deployment

The weights are downloadable under CC-BY-NC-4.0 (weights card); Apache-2.0 (code repo). Check runtime support and size the complete model and working context for your expected concurrency.

Frequently asked questions

Detection, referring localisation, OCR, GUI grounding, keypoints, camera parameters, depth, surface normals, point maps and several kinds of segmentation, all through one instruction interface.

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

briaai

briaai/RMBG-1.4

briaai/RMBG-1.4 is a source-linked Hugging Face repository assigned to Image segmentation. Its recorded task interface accepts image and produces segmentation mask.

— context

briaai

briaai/RMBG-2.0

briaai/RMBG-2.0 is a source-linked Hugging Face repository assigned to Image segmentation. Its recorded task interface accepts image and produces segmentation mask.

— context

CIDAS

CIDAS/clipseg-rd64-refined

CIDAS/clipseg-rd64-refined is a source-linked Hugging Face repository assigned to Image segmentation. Its recorded task interface accepts image and produces segmentation mask.

— context

cocktailpeanut

cocktailpeanut/rm

Hugging Face repository for Image segmentation.

— context