en.Wedoany.com Reported - China's Shishi Technology unveiled its domestic Token optimization factory "Vectron" at the 2026 WAIC exhibition in Shanghai, aiming to enhance the computing efficiency of AI infrastructure. The company's founder and chairman, Yan Bowen, stated at the launch that China has built over 50 intelligent computing centers, with a total intelligent computing power exceeding 1000 EFLOPS. However, there remains significant room for improvement in average utilization efficiency, with substantial computing power being wasted, and poor performance in long-context and multimodal response being common issues.
Yan Bowen noted that total investment in China's computing power industry during the "15th Five-Year Plan" period is expected to reach 7 trillion yuan. If computing power utilization efficiency is improved by 10 to 20 percentage points, it could unlock a trillion-yuan value space. He believes that competition in AI infrastructure is shifting from a "computing power scale race" to a "computing power efficiency race."
The "Vectron" product establishes a comprehensive AI infrastructure optimization system, focusing on addressing issues such as unified scheduling of heterogeneous computing power, deep optimization of large model inference, long-context memory pressure, and adaptation to domestic chip ecosystems. The platform is fully compatible with over 10 types of domestic computing chips, including Ascend, Kunlunxin, Tianshu Zhixin, Taichu, Hanbo Semiconductor, Moore Threads, Muxi, and Enflame, and supports more than 20 mainstream models. It enables unified scheduling and inference acceleration for domestic heterogeneous computing power pools, achieving a daily Token throughput of hundreds of billions.
In terms of core technology, "Vectron" employs result-aware KV Cache compression technology to reduce memory pressure without sacrificing model performance. In full-modal inference scenarios, it uses a training-free Token compression method based on modal adaptation, filtering information for different inputs such as text, video, and speech. For long-context optimization, "Vectron" leverages a mixed position index synthesis method for long-context preference training and a memory-guided re-reading mechanism, achieving performance improvements with significantly less training data than similar methods. Additionally, the product includes a meta-reward-based scalable reward modeling method to support self-optimization of the inference system, as well as a goal-oriented information memory strategy to ensure stable execution of long-term agent tasks.
Founded in 2021, Shishi Technology's founding team originates from the Department of Computer Science at Tsinghua University. Its core members consist of experts in high-performance computing and artificial intelligence, along with scholars from Tsinghua University, Peking University, and Beihang University. The company has served over 200 clients across industries including internet, large model companies, aerospace, biopharmaceuticals, and new energy.
