Pro/zai-org/GLM-5

GLM-5 is the next-generation flagship foundation model launched by Zhipu AI. Its core positioning is as a “general-purpose agent foundation for Agentic Engineering,” specifically designed for complex system engineering and long-term agent tasks. It achieves state-of-the-art performance in coding and agent capabilities among open-source models, with a real programming experience approaching that of Claude Opus 4.5.

• Capability Leap: The model’s parameter scale has expanded from 35.5B (activated at 3.2B) to 74.4B (activated at 4.0B), with pre-training data reaching 28.5T tokens. Combined with the brand-new “Slime” asynchronous reinforcement learning framework and DeepSeek’s sparse attention mechanism, it significantly reduces deployment costs while maintaining context performance around 200K.
• Programming Strength: It scores the highest among open-source models on SWE-bench-Verified (77.8) and Terminal Bench 2.0 (56.2), surpassing Gemini 3.0 Pro. It can independently handle system-level engineering tasks such as backend refactoring, deep debugging, and full-stack development.
• Agent Capabilities: It ranks first among open-source models in multi-tool long-term task evaluations including BrowseComp, MCP-Atlas, and τ²-Bench, demonstrating strong goal alignment, resource scheduling, and multi-step dependency handling capabilities.
• Office Integration: It natively supports GLM in Excel and is compatible with Microsoft’s official AI plugin, enabling spreadsheet automation. It also supports structured outputs (JSON), Function Calls, MCP tool invocations, and context caching.
• Recommended Scenarios: Agentic Coding, end-to-end agent execution, cross-phase office tasks, immersive role-playing, professional script/storyboard generation, precise translation, structured data extraction, and information quality inspection.

───────────────────────────────────────────────────────────────────

Core Capabilities

🧠 Agentic-Ready Architecture

It enables a leap from “writing code” to “writing engineering,” allowing agents to autonomously plan, execute, verify, and deliver complete system-level products.

💻 SOTA-Level Programming Capability

In a real development environment, its experience closely mirrors that of Claude Opus 4.5, significantly reducing manual intervention and covering the entire frontend, backend, and data processing pipeline.

🛠️ Powerful Tool Ecosystem Integration

It supports Function Calls and the MCP protocol, enabling flexible invocation of external tools and data sources to build complex agent workflows.

📊 Deep Adaptation for Office Productivity

Through GLM in Excel and long-context memory, it stably handles multi-step, highly logically connected spreadsheet and document tasks.

🎭 Highly Consistent Role-Playing

In long-text interactions, it consistently maintains character settings, emotions, and narrative logic, delivering evolving, immersive conversations.

📄 Professional Content Generation

Long-text outputs such as scripts and storyboards exhibit production-quality usability, with greatly enhanced character development and plot coherence.

🔍 Precise Information Processing

It can extract structured data from complex texts like contracts, financial reports, and customer service tickets, automatically performing quality checks and risk identification.

⚡ Efficient Inference Optimization

With a 200K context window, a maximum output of 128K, and a sparse attention mechanism, it balances long-term task capabilities with token efficiency.