Zhipu AI Releases 744B-Parameter Open-Source Model
Chinese AI firm Zhipu AI has released GLM-5, an open-source model with a 744B-parameter Mixture-of-Experts (MoE) architecture. The model, trained on 28.5T tokens using Huawei Ascend NPUs, is positioned as a competitor to frontier models like Claude Opus 4.5, demonstrating strong capabilities in agentic reasoning and workflow automation.
- Zhipu AI, known internationally as Z.ai, originated from Tsinghua University's Computer Science Department in 2019 and is considered one of China's "AI Tiger" startups. The company's valuation was reported at $2.8 billion in 2023, with investors including Alibaba, Tencent, Meituan, and Saudi Aramco's Prosperity7 Ventures. - The GLM-5's Mixture-of-Experts (MoE) architecture activates only a fraction of its total 744 billion parameters—roughly 40 billion—for any given token during inference. This approach is designed to provide the knowledge capacity of a very large model while maintaining the computational efficiency and lower inference costs of a smaller model. - A key feature of GLM-5 is its maximum output of 128,000 tokens, a significant increase from the typical 4,000 to 16,000 token limits of many other models. This allows it to generate entire software modules or long-form documents in a single pass. - For enterprise applications, GLM-5 is positioned for "agentic engineering," moving beyond simple code snippet generation to managing entire repositories and complex, multi-step business workflows. Its performance on benchmarks like SWE-Bench and Terminal-Bench 2.0 shows it approaching the capabilities of models like Claude 4.5 Opus in coding and reasoning tasks. - GLM-5's API is priced significantly lower than some competitors, with input tokens costing approximately 80% less and output tokens about 87% less than Claude 4.5 Opus. - The model integrates DeepSeek Sparse Attention (DSA), a technique intended to reduce memory overhead and deployment costs while preserving performance on tasks requiring a long context. It also utilizes a novel asynchronous reinforcement learning framework called "slime" for more efficient post-training. - Zhipu AI has been iterating on its GLM series, with GLM-4.5 released in March 2025 (355B parameters) and GLM-4.7 in December 2025, which showed significant gains in coding and reasoning benchmarks. The company also rebranded internationally from Zhipu AI to Z.ai in 2025. - In January 2025, the United States Commerce Department added Zhipu AI to its Entity List, citing national security concerns.
