SenseTime's Open-Source SenseNova-SI AI Model Said to Outperform GPT-5, Gemini 2.5 Pro in Spatial Intelligence

SenseTime, a partly state-owned Chinese artificial intelligence (AI) company, released its SenseNova-SI series of models on Monday. There are two models, SenseNova-SI-InternVL3-2B and SenseNova-SI-InternVL3-8B, which are now available in open source. The company claims that these models excel in spatial intelligence, making it easy for them to understand how objects move and relate to one another in a three-dimensional plane. Based on internal testing, the company also claims that these models offer higher spatial intelligence than OpenAI's GPT-5 and Google's Gemini 2.5 Pro models.

SenseNova Said to Excel in Spatial Intelligence Benchmarks

The company shared the details about the open-source AI models in a GitHub post. Both the two-billion and eight-billion parameter versions are currently listed for download on Hugging Face. SenseTime is offering these models with a permissive Apache 2.0 license, allowing both research-based usage and commercial use cases.

SenseTime says that despite making significant progress, most multimodal models struggle when it comes to spatial intelligence. Put simply, it is the ability of a large language model (LLM) to understand, reason, and represent how objects or entities relate to one another in space. This capability is important in 3D modelling, navigation and mapping, scene understanding, and even in embodied AI or robots that are powered by AI models.

To solve the problem, the company said it constructed and curated a large-scale, comprehensive collection of spatial intelligence data and trained its models on it. SenseTime added that due to success in building out this model, the company will include SenseNova-SI in larger-scale in-house models as well.

Based on internal testing, the company claimed that the SenseNova-SI-InternVL3-8B scored 62.80, 89.33, and 53.92 in the VSI, MindCube-Tiny, and ViewSpatial benchmarks, respectively. In contrast, all OpenAI's GPT-5, Google's Gemini 2.5 Pro, and xAI's Grok 4 fared worse in these benchmark tests.

More details about these models are not available at this time since the company has not shared the technical report, which typically includes the architecture details and the cooking recipe, including training methods, dataset details, and information about fine-tuning. SenseTime mentioned that the technical report will be shared soon.