China’s AI landscape is exploding with new, powerful large‑language models that promise to make everything from coding to creative work easier and cheaper. Zhipu AI’s latest open‑source model, GLM‑5, shines in long‑range planning and complex system tasks, setting a new benchmark for programming and reasoning. XiYu Technology’s MiniMax M2.5 follows closely, delivering state‑of‑the‑art coding skills and native agent execution with just 10 billion parameters, lowering the barrier for developers. Moonshot AI’s Kimi K2.5 adds a multimodal twist, handling visual programming, parallel agents, and massive text inputs, positioning it as the most versatile open‑source option. Alibaba’s Tongyi Qianwen offers a full suite—language understanding, multimodal interaction, code generation, and AI agents—making it a one‑stop solution for office, R&D, and content creation. DeepSeek focuses on math reasoning, code, and long‑text comprehension, delivering GPT‑4‑level performance at a fraction of the cost. ByteDance’s Doubao 2.0 series (Pro, Lite, Mini, Code) provides a native‑agent architecture that’s both powerful and ultra‑affordable, quickly becoming the most widely adopted model in China. Other notable players include Baidu’s knowledge‑enhanced Wenxin Yiyi, which leverages Baidu’s search and knowledge graph for superior Chinese language understanding, and Tencent’s Hunyuan, which rounds out the ecosystem with strong multimodal capabilities. Together, these models are driving fierce competition, richer AI experiences, and new opportunities for businesses and everyday users alike.
Read moreAI developers are buzzing about agents that can handle long, multi‑step tasks, but the real bottleneck has shifted from raw GPU power to the speed at which large‑language‑model (LLM) caches are read. DeepSeek, together with researchers from Peking University and Tsinghua University, just published a paper that tackles this problem head‑on. Their new "DualPath" dual‑channel loading system cleverly uses otherwise idle bandwidth on storage‑network nodes, feeding the LLM’s cache twice as fast and nearly doubling inference throughput without buying extra GPUs. The breakthrough comes from splitting the pre‑fill and decoding stages into two separate channels and adding a global scheduler that balances traffic, eliminating the network congestion that previously left GPUs idle. The result is a smoother, faster pipeline for agent‑style LLMs that need to keep track of very long conversation histories. DeepSeek also showed that this efficiency gain translates to cheaper training: their R1 model, comparable to GPT‑4, was trained for just $5.57 million, thanks to real‑time gradient monitoring and automatic precision adjustments that keep training speeds soaring at 120 tokens per second on a 14.8‑trillion‑token dataset. In short, smarter software tricks are now outpacing raw hardware upgrades, promising faster, more affordable AI agents for everyone.
Read moreChina’s quantum‑computing community has hit a historic milestone. On Feb. 24, Origin Quantum Computing Technology (Hefei) announced that its home‑grown operating system, **Origin Sinan**, is now freely downloadable from the company’s website. It is the first quantum‑computer OS anywhere in the world that can be obtained, installed and run locally, a move that dramatically lowers the entry barrier for scientists, universities and developers worldwide. Origin Sinan, first released in 2021 and refined through several upgrades, blends “quantum‑superposition‑intelligence” and works with the three dominant hardware approaches – superconducting circuits, ion traps and neutral‑atom platforms. The system ships with one‑click deployment scripts and integrates the QPanda programming framework, letting users write quantum programs and execute them on diverse quantum chips without deep hardware expertise. The open‑source launch is expected to accelerate China’s quantum ecosystem, fostering collaboration and speeding up applications in scientific research and emerging commercial fields. At the same time, the sector is buzzing with investment: startups such as SpinQ, Turing Quantum and Micro‑Epoch have each raised hundreds of millions of yuan in recent financing rounds. Yet challenges remain – quantum hardware is still costly and bulky, and the United States retains a lead of roughly one to two years in raw performance. Nonetheless, Origin Sinan’s release marks a decisive step toward making quantum computing more accessible and could reshape the global race for next‑generation computing power.
Read more