zai-org/glm-4.5

GLM-4.5 is a MoE large model with 355 billion parameters launched by Zhipu AI. Its core positioning is as a full-stack open-source foundation specifically designed for AI agents, natively integrating inference, programming, and tool-call capabilities.

• Leader in Open Source: Achieving an overall score of 63.2 across 12 industry-standard benchmarks, GLM-4.5 ranks first among global open-source models and third globally in overall performance.
• Two Version Configurations: Offering the flagship version GLM-4.5 (355B-A32B) and the lightweight version GLM-4.5-Air (106B-A12B), it meets all-scenario needs from cloud to edge devices.
• Natively Agent-Centric: For the first time, a single model unifies inference, coding, and agent capabilities, enabling it to plan projects, call tools, and execute code just like a human engineer.
• Ultimate Cost-Effectiveness: With extremely low API call costs (as low as 0.8 yuan per million tokens for input) and support for FP8 quantization, it significantly lowers the barrier for enterprise deployment.
• Dual-Mode Switching: Featuring a uniquely developed “Thinking Mode” and “Non-Thinking Mode,” users can freely switch between them based on task complexity, balancing response speed and deep reasoning.

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

Core Capabilities

🤖 Agent Brain: Natively supports tool calls and code execution, enabling it to autonomously complete full-stack development tasks—from requirement analysis to deployment—making it a true digital employee.

⚡ The MoE Efficiency Revolution: With 355 billion parameters, it seamlessly integrates diverse knowledge sources; during inference, only about 9% of parameters (32 billion) are activated, perfectly decoupling high performance from low cost.

💻 Full-Stack Development: With just a single sentence instruction, it can generate complete web pages, games, or databases. It has demonstrated outstanding performance on coding benchmarks such as SWE-bench.

👁️ Visual Reasoning: Combined with GLM-4.5V, it possesses screen understanding and manipulation capabilities, able to recognize icons, locate elements, and even automatically order takeout or book flights for you.

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

Relevant Evaluations

“The New King of Chinese Large Models! GLM-4.5 Tops Open Source Charts, Pitting Logical Reasoning Against Grok 4?”