New Chinese open-source model rivals Claude Opus
Zhipu AI, a Chinese company, has released GLM-5, an open-source frontier model. Early reviews suggest the model is narrowing the performance gap with leading proprietary models like Anthropic's Claude Opus 4.5, particularly in areas like coding and agentic reasoning.
- Zhipu AI, the company behind GLM-5, is a prominent Chinese AI firm founded in 2019 by professors from Tsinghua University. It has attracted significant investment from major tech companies like Alibaba, Tencent, Meituan, and Xiaomi, as well as state-backed funds and international investors like Saudi Aramco's Prosperity7 Ventures. - GLM-5 is a Mixture-of-Experts (MoE) model with a total of 744 billion parameters, with 40 billion active per token. This architecture allows for a large model capacity while maintaining more efficient inference compared to a dense model of the same size. The model was trained on 28.5 trillion tokens. - A notable aspect of GLM-5's development is that it was trained entirely on Huawei Ascend chips, without the use of NVIDIA hardware, representing a significant achievement in China's efforts to build a domestic AI hardware ecosystem. - In benchmark comparisons, GLM-5 shows strong performance, particularly in coding and agentic tasks. On the SWE-bench Verified benchmark, it scored 77.8%, closing the gap with proprietary models like Claude Opus 4.5. On the Vending Bench 2, which measures long-term planning, GLM-5 ranked first among open-source models. - The model is open-sourced under a permissive MIT license, with its weights available on platforms like Hugging Face and ModelScope. This allows for broad accessibility for both research and commercial use. - GLM-5 possesses some unique features, including a context window of up to 200,000 tokens and the ability to generate outputs up to 128,000 tokens in a single pass. It can also directly generate formatted documents such as .docx, .pdf, and .xlsx files from source materials. - While being open-source, deploying GLM-5 requires significant hardware resources, with an estimated 1,490GB of memory needed for BF16 precision. This highlights a growing trend where the accessibility of powerful open-source models is increasingly constrained by hardware availability and cost. - Zhipu AI has been expanding its commercial offerings and partnerships, with use cases in smart customer service, document parsing, and content generation for companies like Deloitte China and Kingsoft Office. The company recently filed for an IPO on the Hong Kong Stock Exchange.