SenseTime's Perspective: Why We Are Firmly Committed to Multimodal General Intelligence

Why multimodal intelligence matters

SenseTime co-founder, executive director, and chief scientist Dahua Lin published a long-form article, Toward Multimodal General Intelligence: SenseTime's Perspective. The article explains why SenseTime regards multimodal general intelligence as a core engine of its technology strategy.

The discussion reviews SenseTime's multimodal journey and outlines the underlying logic, technical path, practical exploration, and future direction of multimodal intelligence.

From visual intelligence to multimodal general intelligence

SenseTime began with computer vision and deep learning, breaking through industrial performance thresholds in face recognition, image quality enhancement, intelligent driving, and other applications. As large language models advanced, SenseTime began to consider how language and vision could meet under scaling laws and create a new generation of intelligent systems.

The company has moved from separate language and vision model tracks toward a unified multimodal model family. Starting from SenseNova 6.0, the previously separated model lines converged into one fused model series, while later versions strengthened interleaved text-image reasoning and multimodal reinforcement learning.

The path to AGI is multimodal

The article argues that intelligence is fundamentally about autonomous interaction with the outside world. Language is a powerful tool for describing the world, but it is not the world itself. To move from language models toward general intelligence, AI must process and integrate information from images, video, audio, spatial signals, and other modalities.

In real applications, value often depends on combining different types of information. Reports contain charts, medical scenarios combine records and images, education uses multimedia materials, and city and industrial scenarios rely heavily on video. Multimodal AI is therefore not only a research path, but also a practical commercial necessity.

Native multimodal training and model design

SenseTime's research compares adaptation training with native training. Adaptation training can add visual capabilities to existing language models at lower cost, but it often struggles to deeply model the intrinsic relationship between language and vision. Native multimodal training, by contrast, integrates modalities during pretraining and allows deeper cross-modal fusion.

The article also discusses data production, model architecture, multimodal reasoning, embodied intelligence, research organization, and how SenseTime balances technological breakthroughs with commercial deployment.

A long-term research direction

SenseTime sees multimodal intelligence as a long-distance race. It requires sustained investment in data, compute, model architecture, evaluation, and productization. The company will continue to push native multimodal models, spatial intelligence, world models, and embodied intelligence toward real-world applications.