Core Insights - SenseNova-SI series models, released by SenseTime, demonstrate superior performance in spatial intelligence benchmarks, particularly the SenseNova-SI-8B model, which achieved an average score of 60.99, significantly outperforming other open-source models like Qwen3-VL-8B (40.16) and BAGEL-7B (35.01) [1][2] - The SenseNova-SI-8B model also surpasses closed-source models such as GPT-5 (49.68) and Gemini-2.5-Pro (48.81) while maintaining the same parameter scale of 8 billion [2] - The performance improvement is attributed to a systematic training design and the establishment of a "spatial capability classification system" by SenseTime, which expanded the scale of spatial understanding data and validated the existence of "scaling law" in this domain [2][5] Model Performance - SenseNova-SI-8B outperformed GPT-5 in various spatial reasoning tasks, showcasing its stability and accuracy in understanding spatial relationships [3][18] - In specific tests, SenseNova-SI-8B consistently provided correct answers while GPT-5 made errors in tasks involving perspective judgment and spatial reasoning [6][10][12][15][16] Technological Advancements - The training methodology for SenseNova-SI incorporates a comprehensive approach to spatial intelligence, categorizing it into six core dimensions: spatial measurement, reconstruction, relationships, perspective transformation, deformation, and reasoning [5] - The model's architecture supports the enhancement of spatial capabilities across various foundational models, indicating a versatile application potential [5] Strategic Implications - The launch of SenseNova-SI aligns with SenseTime's broader strategy in spatial intelligence, complementing their "Wuneng" embodied intelligence platform aimed at improving robots' understanding and adaptability in the physical world [19] - The introduction of the EASI spatial intelligence evaluation platform further supports the development and collaboration within the open-source ecosystem [19] Future Outlook - The ongoing development of spatial intelligence capabilities is crucial for advancing AI's understanding of the physical world, which is essential for applications in autonomous driving and robotics [24]
开源又赢闭源，商汤8B模型空间智能碾压GPT-5，AI看懂世界又进了一步