Realtime AI News
Tsinghua-Affiliated Team Launches Domestic Token Optimization Factory, Compatible with 10+ Chinese Chips
A Tsinghua University-affiliated team has released a domestic Token optimization platform that is compatible with over 10 types of Chinese-made chips and can process 100 billion tokens daily. The solution aims to maximize computing efficiency and provide critical middleware support for China's AI infrastructure.
A Tsinghua University-affiliated team has released a domestically developed Token optimization factory platform that is compatible with over 10 types of Chinese AI chips and capable of processing 100 billion tokens per day, as reported by QbitAI. The product is positioned as middleware infrastructure that converts computing power into usable tokens efficiently.
The platform's core strength lies in its broad compatibility across the Chinese chip ecosystem. As domestic AI chips proliferate but their software ecosystems remain fragmented, a unified Token processing layer that works across multiple chip architectures can significantly reduce developer integration costs.
The optimization factory is purpose-built for large-scale AI inference and training scenarios. A throughput of 100 billion tokens per day means it can handle heavy workloads including batch inference, data processing, and training data preprocessing for large language models.
According to the report, the platform employs advanced compute scheduling and memory management strategies to maximize hardware utilization across different domestic chip platforms. Unlike closed solutions tied to a single chip brand, this open approach lets users choose underlying hardware based on cost, availability, and performance needs.
While Chinese AI chip makers are releasing new products at a rapid pace, the chip itself is only the starting ticket — what determines production efficiency is often the software stack and middleware above it. The Token optimization factory fills a key gap by abstracting fragmented compute resources into a streamlined Token production pipeline.
From an industry perspective, this release signals a shift in China's AI supply chain from hardware breakthroughs toward software ecosystem development. In an environment where GPU supply remains constrained, middleware that can efficiently utilize diverse domestic chips will become a critical component of AI infrastructure competitiveness.
As more Chinese AI chips move into mass production, the value of universal middleware platforms like this Token optimization factory will grow. Key areas to watch include public benchmark data, real-world deployment cases, and whether the platform gets integrated into domestic AI cloud computing platforms.
Why it matters
This platform lowers the adoption barrier for China's domestic AI chip ecosystem and improves compute utilization efficiency, providing practical support for the country's AI infrastructure buildout.
Nearby Updates
All07/18, 11:38
Moonshot AI releases K3 large language model, Elon Musk gives public praise
Moonshot AI, the company behind the Kimi chatbot, has released its next-generation large language model K3, which is claimed to have achieved globally leading performance. Tesla and xAI founder Elon Musk has publicly expressed his appreciation, drawing widespread industry attention.
07/18, 11:41
DeepSeek估值,被一家安徽箱包公司给全部暴露了
DeepSeek估值,被一家安徽箱包公司给全部暴露了. 3500亿估值曝光
07/18, 11:01
Bilibili Becomes WAIC's Official AI Tech Video Platform with 190M+ Monthly AI Content Consumers
Bilibili has become the official AI tech video platform for WAIC 2026, marking its third consecutive year at the conference. The platform reported over 190 million monthly active users consuming AI-related content in Q2 2026, solidifying its position as China's leading AI-focused content community.
07/18, 12:02
Alibaba Cloud Launches AI Agent Security Framework at WAIC 2026 with Three-Layer Unified Protection
Alibaba Cloud released the 2026 Alibaba Cloud AI Agent Security Best Practices at WAIC 2026, introducing a three-layer unified protection system covering infrastructure, model, and application layers. The platform has already served over 300 enterprise clients with more than 500,000 daily API calls.