Key takeaways

• GLM-5 Turbo is specifically designed for fast inference and strong performance in agent-driven environments, optimizing for real-world agent workflows, complex instruction decomposition, and stability across extended tasks.
• Utilizes a Mixture of Experts (MoE) architecture with 744 parameters and supports 'text+image->text' modality, allowing processing of both text and images for text output.
• Excels in agent workflows with improved instruction decomposition, tool use, scheduled execution, and stability, boasting a large context length of 202,800 and max output tokens of 120,000.
• Achieved 96.9% accuracy on the MATH benchmark subset, demonstrating strong mathematical problem-solving skills, particularly in Algebra and Number Theory, with detailed step-by-step explanations.
• Struggles with complex combinatorial and intricate geometry problems, occasional errors in interpreting problem constraints, and minor calculation mistakes in multi-step problems.

Technical specifications

Context length

203k tokens

Max output

—

Architecture

MoE

Modalities

text+image->text

Parameters

~744B

Open weights

No

Licence

Proprietary

Regions

China

Vendor

z-ai

Provider model key

z-ai/glm-5-turbo