Model Releases
Zhipu AI Unveils GLM-4.5: The 355B MoE Open Source Powerhouse
Zhipu AI releases GLM-4.5, a 355B parameter MoE model offering superior reasoning and coding capabilities at a fraction of the cost of competitors like DeepSeek.
July 28, 2025
Model ReleaseGLM-4.5
Introduction
In a significant move for the open-source AI community, Zhipu AI has officially released GLM-4.5, marking a new era in large language model efficiency and capability. Released on July 28, 2025, this flagship model addresses the growing demand for high-performance, cost-effective inference solutions. As developers seek alternatives to expensive proprietary APIs, GLM-4.5 positions itself as a robust open-source contender that balances massive parameter counts with architectural efficiency.
The release comes amidst a period of rapid innovation in the Chinese AI sector, where Zhipu AI has reported a 132% rise in annual revenue driven by AI adoption. GLM-4.5 is not just an incremental update; it represents a strategic shift towards specialized MoE (Mixture of Experts) architectures that reduce inference costs without sacrificing performance. For engineering teams looking to integrate advanced reasoning and coding assistants into production pipelines, this model offers a compelling technical foundation.
- Released: 2025-07-28
- Provider: Zhipu AI
- License: Open Source (Apache 2.0)
- Architecture: Mixture of Experts
Key Features & Architecture
GLM-4.5 is built on a 355B parameter MoE architecture, which allows the model to dynamically activate specific expert sub-networks based on the input task. This design significantly reduces the computational load during inference compared to dense models of similar size. The model supports a massive context window of 128,000 tokens, enabling it to process long documents, codebases, and complex data streams with ease. Furthermore, it features native multimodal capabilities, allowing for the ingestion of text, code, and structured data simultaneously.
The model is optimized for agentic workflows, meaning it can plan, execute, and refine tasks autonomously. This is a critical feature for enterprise applications where AI agents need to interact with internal tools and databases. Zhipu claims the model is significantly cheaper to run than DeepSeek's V3 series, primarily due to its sparse activation rates and optimized tokenization strategies. Developers can expect high throughput on standard GPU clusters without requiring specialized hardware.
- Parameters: 355B Total (MoE)
- Context Window: 128k Tokens
- Multimodal: Text, Code, Structured Data
- Agentic: Native Planning & Tool Use
Performance & Benchmarks
In independent benchmarking, GLM-4.5 demonstrates strong reasoning and coding proficiency. On the MMLU (Massive Multitask Language Understanding) benchmark, it scores 88.5%, outperforming several previous open-source generations. For developers, the HumanEval benchmark score of 92.1% highlights its capability in generating syntactically correct and functional code snippets. These metrics suggest that GLM-4.5 is ready for integration into software development workflows where accuracy is paramount.