Key takeaways

• GLM 4.5 Air is a lightweight, agent-centric Mixture-of-Experts (MoE) model designed for real-time interaction and advanced reasoning in agent-centric applications. It supports hybrid inference with a 'thinking mode' for complex reasoning/tool use and a 'non-thinking mode' for real-time interaction.
• The model demonstrates strong mathematical reasoning (92.59% accuracy on AIME 2024), particularly excelling in algebraic manipulation, geometric formulas, and accurate arithmetic, but struggles with complex multi-step reasoning and creative problem-solving.
• GLM 4.5 Air's performance on the 'Humanity's Last Exam' dataset was poor (4.36% accuracy), indicating difficulties with complex, multi-step problems, factual recall across diverse domains, and nuanced interpretation.
• The model exhibits a discrepancy between its high confidence scores and its actual accuracy, particularly in 'Physics' and numerical 'Biology/Medicine' tasks, often showing high confidence in incorrect answers.
• Known limitations include struggles with intricate function definitions, advanced number theory, image-based questions, and tasks requiring external world knowledge not explicitly provided in the prompt.

Technical specifications

Context length

128k tokens

Max output

—

Architecture

MoE

Modalities

text->text

Parameters

~106B

Open weights

Yes

Licence

MIT

Regions

China

Vendor

z-ai

Provider model key

z-ai/glm-4.5-air