Compared with GLM-4.5, this generation brings several key improvements:
Longer context window: The context window has been expanded from 128K to 200K tokens, enabling the model to handle more complex agentic tasks.
Superior coding performance: The model achieves higher scores on code benchmarks and demonstrates better real-world performance in applications such as Claude Code、Cline、Roo Code and Kilo Code, including improvements in generating visually polished front-end pages.
Advanced reasoning: GLM-4.6 shows a clear improvement in reasoning performance and supports tool use during inference, leading to stronger overall capability.
More capable agents: GLM-4.6 exhibits stronger performance in tool using and search-based agents, and integrates more effectively within agent frameworks.
Refined writing: Better aligns with human preferences in style and readability, and performs more naturally in role-playing scenarios.
Evaluations
16
across 14 benchmarks
Latency
22m 47s
Context length
203k
tokens
Cost
$0.60 · $2.20
input · output per 1M tokens
Key takeaways
GLM 4.6 is a text generation model designed for complex agentic tasks, coding, and general text processing, expanding on GLM-4.5 with a longer context window, enhanced coding, and refined reasoning.
The model utilizes a Mixture of Experts (MoE) architecture with a 200K token context window. Specific training details beyond being a refined version of GLM-4.5 are not provided.
GLM 4.6 excels in coding, advanced reasoning with tool use, and strong integration within agent frameworks, and provides refined, human-preference-aligned writing. However, it struggles with biology and scientific reasoning, shows calculation inaccuracies, and exhibits overconfidence in incorrect answers.
It demonstrated strong logical reasoning and computational prowess for general mathematical problems, achieving 96.67% accuracy on the 'MATH-500' dataset.
Conversely, it scored only 10.21% accuracy on the 'Humanity's Last Exam' dataset (Biology/Medicine Benchmark), indicating significant weaknesses in applying reasoning and calculation to biology and medicine-related questions, often coupled with overconfidence.
