Key takeaways

• GLM 4.7 is Z.AI's flagship 'text+image-to-text' model, built with a Mixture of Experts (MoE) architecture and 358 parameters, designed for complex agent tasks with enhanced programming capabilities and stable multi-step reasoning, aiming for natural conversational experiences.
• The model demonstrates strong capabilities in solving AIME-level mathematics problems, achieving 96.67% accuracy on the AIME 2024 dataset, showcasing high proficiency in number theory, combinatorics, and geometry with clear and structured derivations.
• GLM 4.7 struggles significantly with precise numerical computation and visual data interpretation, scoring zero on image-based tasks. It also performs poorly in interactive web-based environments with only a 17.82% success rate, due to numerical errors, misinterpretation of complex constraints, inability to process image-based data, or recall niche facts.
• The model completely fails on all image-based tasks, highlighting a critical area for development in multimodal understanding, especially for tasks requiring visual input processing.
• Addressing precision issues and enabling visual input processing are essential steps to improve GLM 4.7's utility in specialized domains, as its complete failure on image-based tasks highlights a critical area for development in multimodal understanding.

Technical specifications

Context length

200k tokens

Max output

—

Architecture

MoE

Modalities

text+image->text

Parameters

~358B

Open weights

Yes

Licence

MIT

Regions

China

Vendor

z-ai

Provider model key

z-ai/glm-4.7