SenseTime released SenseNova-SI-1.3, which ranks first on EASI-8 and improves performance on tasks.
AI software company SenseTime has open-sourced SenseNova-SI-1.3, a spatial intelligence model that delivers improvements across tasks, including metric measurement, perspective-taking, and broader reasoning.
On EASI, an evaluation platform that integrates multiple spatial intelligence benchmarks, SenseNova-SI-1.3 achieved the top average score on EASI-8, a unified evaluation across eight benchmarks, and surpassed Gemini-3-Pro overall.
SenseTime described EASI-8 as deliberately challenging and said it includes tasks that often prove difficult for models such as Gemini-3-Pro. Examples include:
- Counting building models correctly across two viewpoints by matching objects between images and avoiding double-counting
- Inferring the orientation of a study area by recognising that two partial photos show the same room
- Determining left/right directions from another person’s perspective rather than the viewer’s viewpoint
- Using multiple images to infer where an object sits relative to a bottle from a specific view
- Judging the correct direction of a bus stop from the visual scene rather than relying on assumptions
The release cited a 2025 ICML paper, Core Knowledge Deficits in Multi-Modal Language Models, which it said found perspective transformation has low correlation with other multimodal capabilities and that increasing model size does not necessarily improve perspective-taking performance.
With larger volumes of spatial intelligence data, its 8B-parameter SenseNova-SI base model surpassed closed-source models such as GPT-5 on perspective-taking.
Stay updated on crypto and AI by following our socials
