Recently, SenseTime has made significant progress in the field of spatial intelligence and officially released and open-sourced the SenseNova-SI series of models. This series of models has performed outstandingly in multiple authoritative evaluations, not only surpassing similar open-source multimodal models but also exceeding international top closed-source models such as GPT-5 and Gemini 2.5Pro.
Currently, mainstream large models perform well in tasks such as knowledge, writing, and reasoning, but they show obvious shortcomings in spatial understanding and reasoning. SenseTime conducted in-depth research on this issue and found that spatial intelligence is a key capability for embodied agents to interact with the real world. Therefore, SenseTime launched the SenseNova-SI model series to improve AI's performance in spatial intelligence.
This series of models includes two versions: 2B and 8B. The latest evaluation results show that the SenseNova-SI-8B model achieved an average score of 60.99 in spatial intelligence benchmark tests, significantly outperforming multiple models such as Qwen3-VL-8B and BAGEL-7B, and also surpassing top closed-source models like GPT-5 and Gemini-2.5Pro. This breakthrough by SenseTime is not just a simple performance improvement, but a qualitative leap.
SenseTime's success comes from its systematic training method, which enhances the model's spatial understanding capabilities through large-scale high-quality datasets, validating the effectiveness of the "scale effect." This training paradigm is applicable to different base model architectures and can achieve a comprehensive enhancement of capabilities in six core dimensions: spatial measurement, spatial reconstruction, spatial relationships, perspective transformation, spatial deformation, and spatial reasoning.
In addition, SenseTime also released the EASI spatial intelligence evaluation platform, aiming to promote the standardization of spatial intelligence technology assessment and provide reliable evaluation benchmarks for academia and industry. The release of SenseNova-SI will lay a solid foundation for AI applications in fields such as autonomous driving and robotics.
EASI Spatial Intelligence Evaluation Platform: https://github.com/EvolvingLMMs-Lab/EASI
Key Points:
🌟 SenseTime launched the SenseNova-SI model series, with spatial intelligence performance leading international top models.
📊 The SenseNova-SI-8B achieved a high score of 60.99 in spatial intelligence evaluation, significantly outperforming other similar models.
🔍 SenseTime's systematic training method improves the model's spatial understanding capabilities through large-scale data.
