Technical Specifications

System Requirements

VRAM requirements for different quantization methods and context sizes

GLM-4.5

The GLM-4.5 model, developed by Z.ai (formerly Zhipu AI), represents their latest flagship hybrid reasoning model, designed to unify reasoning, coding, and agentic capabilities within a single architecture. This model is specifically optimized for agent-oriented applications, providing advanced functionalities for complex problem-solving. It is offered alongside a lighter variant, GLM-4.5-Air, which is optimized for efficiency while retaining core capabilities.

Architecturally, GLM-4.5 leverages a Mixture-of-Experts (MoE) design. It features a total of 355 billion parameters, with 32 billion active parameters utilized during a forward pass, aiming for higher parameter efficiency compared to other models. The model supports a dual reasoning approach, incorporating a "Thinking Mode" for intricate reasoning, multi-step planning, and tool usage, and a "Non-Thinking Mode" for rapid, instantaneous responses. This hybrid approach allows for flexibility in deployment, accommodating both deep analytical tasks and low-latency interactive scenarios.

GLM-4.5 is engineered for robust performance in domains such as tool invocation, web browsing, and software engineering, including both frontend and backend development. It supports native function calling and can be integrated into code-centric agents. The training regimen for GLM-4.5 involved an initial pretraining phase on 15 trillion tokens of general-domain data, followed by fine-tuning on an additional 7 trillion tokens focused on code and reasoning datasets. Reinforcement learning, specifically using Z.ai's custom-built 'slime' engine, was applied to further enhance its reasoning, coding, and agentic capabilities. The model is designed to handle extended conversational contexts, supporting a context length of 128,000 tokens and a maximum output token limit of 96,000 tokens.