On December 15, SENSETIME-W (00020) officially launched Seko 2.0—the industry's first multi-episode generative AI agent—leveraging its technological expertise in generative AI and multimodal interaction. This agent demonstrates significant advantages in maintaining consistency across multi-episode video generation, powered by SENSETIME's self-developed Rixin Seko series models, including SekoIDX and SekoTalk, which form the foundation for image and video generation.
The Rixin Seko series models have now been successfully adapted to domestic AI chip manufacturer Cambricon Technologies Corporation Limited (688256.SH), marking a critical leap in domestic computing power's support for AIGC core scenarios—from language to multimodal capabilities. This advancement not only deepens technical collaboration but also strengthens China's AI ecosystem, providing more robust and independent infrastructure for visual content innovation.
SENSETIME's LightX2V framework features a highly compatible domestic adaptation plugin mode, enabling rapid integration with various domestic hardware, including Cambricon chips. To maximize domestic computing power efficiency, the Seko series models and LightX2V framework incorporate hardware-friendly innovations such as low-bit quantization, compressed communication, and sparse attention mechanisms, improving inference performance by over 3x.
In October, SENSETIME and Cambricon entered a strategic partnership to jointly optimize software and hardware while fostering an open, collaborative industrial ecosystem. Their latest achievement in adapting multimodal generative models represents a major milestone in domestic large-scale model and computing power synergy, allowing developers and enterprises to access cutting-edge multimodal AI capabilities at lower costs.
Post-adaptation, the two companies will further optimize in multiple areas: 1. **Enhancing Core Model Capabilities**: Improving long-sequence processing and low-bit computing to boost efficiency and response speed without compromising model performance. 2. **Increasing Computing Efficiency & Cost-Effectiveness**: Leveraging operator fusion and auto-tuning to reduce resource consumption, making high-performance multimodal AI more accessible. 3. **Scaling Parallel Processing**: Optimizing cross-hardware scheduling and communication strategies to enhance efficiency and stability in large-scale cluster tasks. 4. **Flexible Resource Management**: Exploring hierarchical scheduling and heterogeneous resource coordination to minimize memory pressure and ensure stable operation across diverse environments.
Through these efforts, SENSETIME and Cambricon aim to elevate model efficiency, computing power utilization, and cross-hardware compatibility, lowering barriers to multimodal AI adoption while improving user experience. Together, they will drive the growth of China's AI application ecosystem, delivering efficient, user-friendly solutions and fostering an open, developer-friendly environment to spur innovation.
Comments