Zhipu AI Unveils GLM-5 Model Trained Without US Chips
Zhipu AI has announced its GLM-5 model, which it claims rivals international counterparts like Claude and GPT-5 in reasoning and agentic capabilities. The model was reportedly trained entirely on a China-based hardware stack, demonstrating technological independence from US export controls. The company is now integrating GLM-5 into consumer-facing agent platforms, emphasizing its efficiency on domestic chips.
- Zhipu AI's GLM-5 is a Mixture-of-Experts (MoE) model with 744 billion total parameters, with 40 billion active during inference, and was trained on 28.5 trillion tokens. It supports a 200k token context window and utilizes DeepSeek Sparse Attention for efficiency in processing long sequences. - The company has received significant funding from major Chinese tech firms and state-backed investors, including Alibaba, Tencent, Meituan, and various government funds from cities like Beijing, Shanghai, and Hangzhou, accumulating over $1.4 billion in total. - GLM-5's development on domestic hardware, including Huawei's Ascend chips, is part of a broader national initiative in China to achieve self-sufficiency in AI and semiconductor technology, a goal emphasized by President Xi Jinping. This push includes substantial government investment and tax incentives for local companies like Huawei, Baidu, and Alibaba to close the gap with global competitors. - For orchestrating multiple agents, frameworks like CrewAI, LangGraph, and Microsoft's AutoGen are gaining traction. Architectural patterns being adopted include sequential orchestration for linear workflows, concurrent execution for parallel tasks, and more dynamic, manager-led patterns where a central agent delegates tasks to specialized agents. - Key challenges in scaling multi-agent systems for consumer applications include ensuring reliable, deterministic workflows, managing complex integrations with legacy systems, and maintaining performance as user volume and input diversity grow. Many development teams are adopting modular, API-first architectures and message-based patterns to improve scalability and reliability. - Zhipu AI is one of China's "AI Tigers," a group of prominent AI startups that also includes DeepSeek, Moonshot, Minimax, 01.AI, and Baichuan, all competing to advance China's standing in the global AI race. The competitive landscape is heating up, with rivals like DeepSeek also pioneering new efficient architectures and ByteDance releasing advanced video generation models. - The GLM-5 model is being positioned for "agentic engineering," focusing on complex, long-horizon tasks like multi-step coding and reasoning. It is accessible via Zhipu's Z.ai chat platform and APIs, and the model weights have been released on open-source platforms like Hugging Face, with an MIT license expected, allowing for commercial use.
