How SenseTime’s New Multimodal Architecture Redefines Unified AI Foundations

At the 2026 World Artificial Intelligence Conference, the spotlight turned to embodied intelligence and AI terminals, while a parallel technical track highlighted the evolution of multimodal model architectures.

Traditional multimodal systems suffer from a "building‑blocks" approach where understanding, generation, editing, and action are split across specialist modules, leading to severe information decay as task chains lengthen. The industry now seeks new solutions.

SenseTime introduced two complementary models: SenseNova U1 Pro, a delivery‑grade native multimodal agent base for long‑range tasks, and the open‑source SenseNova‑Vision visual foundation model. U1 Pro integrates perception, generation, and action within a unified capability framework.

The core of U1 Pro is the NEO‑unify architecture, which shares a single representation space for understanding, generation, and editing. Its design includes a near‑lossless visual interface, a native Mixture‑of‑Transformer (MoT) that jointly handles vision and language, and a unified learning scheme that trains text with autoregressive cross‑entropy and visual data with pixel‑flow matching.

U1 Pro demonstrates several practical advantages: (1) superior design‑aesthetic reasoning that captures layout, color, and hierarchy for professional‑grade graphics; (2) native 8K ultra‑high‑resolution output that preserves text and icon fidelity at large scales; (3) fine‑grained control over image‑text details, coupling generation with precise information expression; and (4) a long‑term agentic closed‑loop that performs dozens of generation‑check‑revise cycles, enabling end‑to‑end visual delivery such as automated sports‑analysis reports.

SenseNova‑Vision tackles the classic vision stack—detection, segmentation, depth, 3‑D reconstruction—by reformulating all tasks as a native multimodal generation problem, eliminating task‑specific heads. It encodes categories, coordinates, OCR results as symbolic answers via text generation, dense predictions (masks, depth, normals) via image generation, and complex scenes through mixed image‑text output, preserving both modalities.

To support this "unified" paradigm, SenseTime released the SN‑VC‑50M dataset, containing 50 million high‑quality visual supervision samples with heterogeneous annotations (boxes, keypoints, OCR, masks, depth, normals, point clouds, camera poses) converted to a standard "visual input + natural‑language instruction + decodable response" format.

Benchmark results (arXiv:2607.06560) show SenseNova‑Vision achieving state‑of‑the‑art performance on structured visual understanding while matching top expert models on dense geometry, segmentation, and multi‑view 3‑D tasks. The unified training also enables cross‑task capability recombination, allowing the model to handle unseen task combinations without module switching.

Looking ahead, SenseTime plans to fuse SenseNova‑Vision’s physical‑world perception and 3‑D reconstruction abilities into the U series, creating a "full‑perception, full‑generation" multimodal base capable of long‑range creation, high‑density image‑text understanding, and spatial modeling—positioned as a core backbone for next‑generation AGI.