GLM-4.5

GLM-4.5 is a cutting-edge open-weight Mixture of Experts (MoE) model, boasting 355 billion total parameters with 32 billion active parameters. Designed to excel in reasoning, coding, and agentic tasks, it represents a significant step toward unifying diverse capabilities within a single framework. The model is complemented by GLM-4.5-Air, a lighter variant with 106 billion total parameters and 12 billion active parameters. Both models support dual-mode inference, offering a "thinking mode" for complex reasoning and tool usage, and a "non-thinking mode" for instant responses.

GLM-4.5 achieves strong performance across agentic, reasoning, and coding benchmarks, ranking third overall among leading models from OpenAI, Anthropic, Google DeepMind, and others. For agentic tasks, it offers an extended 128k context length and native function-calling capabilities, excelling in benchmarks like τ-bench and BFCL-v3. It demonstrates superior efficiency in web browsing applications, outperforming competitors like Claude-4-Opus on BrowseComp, a challenging benchmark for multi-turn tool usage and reasoning.

In reasoning, GLM-4.5 leverages its deep architecture and advanced attention mechanisms to tackle complex problems in mathematics, science, and logic. It performs exceptionally well on benchmarks such as MMLU and AIME, showcasing robust reasoning and problem-solving capabilities. The model also excels in coding tasks, seamlessly integrating with coding toolkits like Claude Code and CodeGeex. It supports full-stack development, enabling the creation of sophisticated web applications and standalone artifacts, such as interactive mini-games and physics simulations.

GLM-4.5 also demonstrates advanced agentic coding capabilities, achieving high success rates in multi-round human interactions and tool usage scenarios. It leads in tool-calling reliability, with a 90.6% success rate, outperforming models like Claude-4-Sonnet and Qwen3-Coder. Its ability to autonomously generate presentation materials, such as slides and posters, further underscores its versatility.

The model's architecture incorporates innovations like loss-free balance routing, grouped-query attention, and the Muon optimizer to enhance efficiency and reasoning capacity. Pre-training spans general and domain-specific datasets, followed by reinforcement learning (RL) to refine agentic and reasoning capabilities. The RL stage employs a hybrid training architecture via the open-sourced slime framework, designed for scalability and efficiency in complex tasks.

GLM-4.5 is available through Z.ai, the Z.ai API, and open-weight repositories on HuggingFace and ModelScope, supporting local deployment and integration with coding agents. Its comprehensive capabilities make it a powerful tool for addressing diverse challenges in reasoning, coding, and agentic applications.