SenseTime’s SenseNova Seko Series Models Successfully Adapted to Cambricon: A Major Advancement in Chinese Computing Power and Multimodal AI

SenseTime officially launched Seko 2.0 — the industry's first multi-episode generation agent, leveraging SenseTime’s technical accumulation in generative AI and multimodal interaction. This agent has demonstrated significant advantages in the consistency of multi-episode video generation, backed by a technical foundation built on SenseTime's self-developed SenseNova Seko series models, including multimodal image and video generation models, such as SekoIDX and SekoTalk.

The successful adaptation of the SenseNova Seko series models to Cambricon marks a key leap in Chinese computing power support for core AIGC scenarios, advancing from language-centric to multimodal capabilities. In addition to deepening technical synergy, this achievement represents a substantial improvement in the domestic AI ecosystem, providing more solid and independent underlying support for the innovative development of visual content.

Notably, SenseTime's LightX2V framework features a highly compatible domestic adaptation plugin mode, which facilitates rapid integration with various Chinese hardware. Currently, it supports multiple Chinese chips including Cambricon. To maximize the potential of domestic computing power, the Seko series models and LightX2V framework were designed with hardware-friendly innovative mechanisms — such as low-bit quantization, compressed communication, and sparse attention —  improving inference performance by over three times.

In October this year, SenseTime and Cambricon established a strategic partnership, focusing on advancing the joint optimization of software and hardware and building an open, win-win industrial ecosystem. The recent adaptation of multimodal generative models to Cambricon represents a crucial phase in the collaborative innovation between Chinese large models and Chinese computing power bases, allowing more developers and enterprises to access top-tier multimodal AI capabilities at lower costs.

Following the completion of the adaptation, SenseTime and Cambricon will pursue in-depth optimization in multiple directions:

• Continuously optimizing core model capabilities: We will continue to refine long-sequence processing, low-bit computing, and other key areas, further improving the overall efficiency and response speed of multimodal generation while maintaining model  performance.

• Improving computing power utilization and cost efficiency: Through technologies like operator fusion and automatic operator optimization, we will promote more efficient computing and storage methods, minimize resource consumption during model operations,  and enable enterprises to obtain high-performance multimodal capabilities at lower costs.

• Strengthening large-scale parallel processing capabilities: Utilizing optimization technologies such as computation-communication parallelism, we will enhance cross-hardware scheduling and communication strategies. This will improve operational    efficiency and stability for complex tasks in large-scale clusters.

• Building a more flexible resource management mechanism: We will explore hierarchical scheduling and heterogeneous resource coordination technologies to effectively reduce video memory pressure while maintaining performance, enabling models to run stably in a wider range of environments.

Through comprehensive collaborative optimization between SenseTime and Cambricon, we aim to comprehensively enhance model efficiency, significantly improve computing power and resource utilization, and optimize collaborative adaptation across hardware environments. This will further lower the threshold for multimodal AI adoption and enhance the overall user experience.

SenseTime and Cambricon will jointly promote the growth and development of the domestic AI application ecosystem. We aim to provide a more refined, efficient and user-friendly stepped product system, and build more open, developer-friendly tools and ecosystems to stimulate the innovation potential of cutting-edge applications.