SenseNova-SI — SenseTime’s open-source Spatial Intelligence large model

What is SenseNova-SI?

SenseNova-SI is SenseTime’s open-source spatial intelligence large model, designed specifically to enhance spatial understanding. Trained on large-scale, high-quality spatial datasets, the model significantly improves its capabilities in core dimensions such as spatial measurement, relational understanding, and viewpoint transformation.
In multiple authoritative benchmarks, SenseNova-SI outperforms open-source models of similar size and even surpasses top-tier closed-source models like GPT-5. The project provides detailed installation and usage guides to help developers get started quickly, advancing embodied AI and world-model research while laying the foundation for AI systems that can understand the 3D world.

Key Features of SenseNova-SI

• Spatial measurement and estimation:
  Accurately estimates object dimensions, distances, and other quantitative spatial attributes.

• Spatial relationship understanding:
  Understands relative positions, orientations, and spatial layouts between objects.

• Viewpoint transformation:
  Handles changes in scene appearance from different viewpoints and infers the effects of viewpoint shifts.

• Spatial reconstruction and deformation:
  Understands 3D structures of objects and maintains spatial awareness after deformation or reconstruction.

• Spatial reasoning:
  Performs logical reasoning based on spatial information, such as predicting object movement or layout changes.

• Multimodal fusion:
  Integrates image, text, and other modalities to better comprehend complex spatial scenes.

Technical Principles of SenseNova-SI

• Scale effect:
  Through training on massive amounts of high-quality spatial data, SenseTime validates the “scale effect”—increased data volume leads to significant improvements in spatial cognition. This is the core driver behind SenseNova-SI’s performance jump.

• Systematic training methodology:
  SenseTime defines a classification framework for spatial capabilities and expands the dataset based on it, applying a systematic training approach that improves all dimensions of spatial intelligence in a consistent manner.

• Multimodal fusion architecture:
  Built on architectures such as InternVL, SenseNova-SI effectively fuses visual and textual information to enhance understanding of complex scenes.

Project Links

• GitHub repository: https://github.com/OpenSenseNova/SenseNova-SI

• HuggingFace model collection: https://huggingface.co/collections/sensenova/sensenova-si

Application Scenarios of SenseNova-SI

• Autonomous driving:
  With accurate spatial measurement and viewpoint transformation, vehicles can better interpret road environments, predict object movement, and improve safety and reliability.

• Robotic navigation and manipulation:
  Spatial relationship understanding and reasoning enable robots to navigate complex environments and operate objects with greater precision.

• Virtual reality and augmented reality:
  Enhances the realism of virtual spaces and enables more natural user interaction in VR/AR environments.

• Intelligent security systems:
  Analyses surveillance footage with spatial intelligence to quickly detect anomalies or changes in object positions, improving the efficiency and accuracy of security monitoring.

• Architecture and spatial planning:
  Assists designers with 3D spatial layout planning and rapidly generates or optimizes design schemes using spatial reconstruction capabilities.