GLM-5: a flagship LLM series for long-horizon agentic engineering and complex systems tasks

What it solves

GLM-5 is a series of flagship large language models designed for complex systems engineering and long-horizon agentic tasks. It addresses the limitation of previous models that plateau in performance when given more time or iterations, instead sustaining productivity over hundreds of rounds of tool calls and thousands of iterations to solve ambiguous problems.

How it works

The series evolves through three versions:

• GLM-5: Scales to 744B parameters (40B active) and uses DeepSeek Sparse Attention (DSA) to reduce deployment costs. It was trained using a novel asynchronous RL infrastructure called "slime" to improve training throughput.
• GLM-5.1: Focuses on agentic engineering and coding, improving the model's ability to break down complex problems, run experiments, and revise strategies through repeated iteration.
• GLM-5.2: Introduces a solid 1M-token context window and a new architecture called IndexShare, which reuses the same indexer across sparse attention layers to reduce per-token FLOPs by 2.9x at long context lengths. It also features an improved MTP layer for speculative decoding.

Who it’s for

This project is for developers and researchers working on autonomous agents, complex software engineering tasks, and applications requiring the processing of massive contexts (up to 1M tokens).

Highlights

• Long-Horizon Capability: Specifically optimized for tasks that require long-term planning and resource management.
• 1M Token Context: GLM-5.2 provides a stable, large context window for long-horizon work.
• Flexible Thinking Effort: Supports a reasoning_effort parameter (max or high) to balance performance and latency.
• High-Performance Coding: Outperforms many open-source models on coding benchmarks like Terminal-Bench and SWE-bench Pro.