Key takeaways

• GLM-5 is an open-source, text+image-to-text foundation model by Z.ai, specifically designed for complex systems design and long-horizon agent workflows in large-scale programming tasks, targeting expert developers.
• The model utilizes a Mixture of Experts (MoE) architecture to deliver production-grade performance, incorporating advanced agentic planning, deep backend reasoning, and iterative self-correction.
• GLM 5 excels in complex arithmetic, algebraic manipulations, and geometry problems, consistently providing detailed and correct step-by-step solutions (e.g., 97.2% accuracy on MATH-500, 98.32% on 'Truthfulness Benchmark'). It also shows foundational reasoning in physics and some biology.
• Key limitations include struggles with numerical precision in complex multi-step scientific calculations, difficulty with counterintuitive conclusions, inconsistent performance in image-based questions, and occasional misinterpretation of complex phrasing in science benchmarks.
• The model performs well in tasks requiring mathematical accuracy, but struggles with nuanced, domain-specific knowledge, visual inputs, and strict output formatting, even when numerical values are correct (e.g., 10.37% accuracy on 'Humanity's Last Exam').

Technical specifications

Context length

203k tokens

Max output

—

Architecture

MoE

Modalities

text+image->text

Parameters

~744B

Open weights

Yes

Licence

MIT

Regions

China

Vendor

z-ai

Provider model key

z-ai/glm-5