SenseTime Ditches VAE/VE in Multimodal Overhaul

🔑 Enhanced Key Takeaways

• •NEO architecture, open-sourced in collaboration with Nanyang Technological University's S-Lab, innovates in attention mechanisms, positional encoding, and semantic mapping for native vision-language fusion[1][2][4][7].
• •NEO-based models in 2B and 9B sizes achieve top performance with 1/10th the data via Cross-View Prediction training, surpassing GPT-5 and Gemini-3 Pro in spatial intelligence[1][3][4].
• •SenseNova V6.5 ranked No.1 in China for multimodal tasks including facial recognition, 3D object recognition, and medical image analysis in 2025 evaluations[1].

🛠️ Technical Deep Dive

• •NEO uses Pre-Buffer & Post-LLM dual-stage integration to preserve LLM language reasoning while building visual capabilities, avoiding degradation in cross-modal training[2][4].
• •Innovations include unified processing of visual and language data at core architecture level, enabling end-to-end native integration for robotics, video understanding, and 3D interaction[2][4][7].
• •SenseNova-SI leverages spatial capability classification and large-scale diverse data, proving scaling laws in six dimensions: metric measurement, mental reconstruction, spatial relationships, perspective-taking, deformation/assembling, and reasoning[3].
• •Open-sourced NEO models: 2B and 9B specifications, with technical report at arxiv.org/abs/2511.13719 detailing spatial intelligence advancements[3][4][7].

🔮 Future ImplicationsAI analysis grounded in cited sources

⏳ Timeline