Key takeaways

• GLM-4.5 is an MoE foundation model designed for agent-based applications, excelling in reasoning and code generation for complex problem-solving. It supports a context length of 128k tokens.
• The model employs a hybrid inference mode with both a 'thinking mode' for complex reasoning/tool use and a 'non-thinking mode' for instant responses, offering user control over reasoning behavior.
• GLM-4.5 demonstrates strong mathematical reasoning capabilities, achieving 91.8% accuracy on the 'MATH-500' dataset.
• Despite its strengths, GLM-4.5 has limitations in instruction-following, readability, and strict length constraints, and occasionally exhibits hallucination and toxicity.
• The model performed poorly on 'Humanity's Last Exam' with only 4.1% accuracy and 48.28% readability, indicating challenges in broad knowledge, complex problem-solving, and adhering to instructions for less defined tasks.

Technical specifications

Context length

128k tokens

Max output

—

Architecture

MoE

Modalities

text->text

Parameters

~358B

Open weights

Yes

Licence

MIT

Regions

China

Vendor

z-ai

Provider model key

z-ai/glm-4.5