Z.AI Launches GLM-5.1: A Record-Breaking Agentic Model

Z.AI has announced the release of its new model, GLM-5.1, designed specifically for agentic tasks and boasting significantly enhanced coding capabilities compared to its predecessor. The model has achieved outstanding results on SWE-Bench Pro, far surpassing GLM-5 in repository generation and real-world terminal tasks.

The architecture of GLM-5.1 is based on DSA and MoE, allowing for a significant reduction in training and inference costs while maintaining high accuracy with long contexts. It employs a glm_moe_dsa architecture, making it more efficient compared to standard dense models, which is crucial for developers considering self-hosting the model.

A key innovation is the asynchronous reinforcement learning infrastructure, which greatly enhances the model's post-training efficiency, enabling it to learn from complex interactions. GLM-5.1 is capable of sustaining high productivity over longer sessions, making it more effective in tackling complex tasks.

On SWE-Bench Pro, GLM-5.1 achieved a score of 58.4, outperforming competitors like GPT-5.4 and Claude Opus 4.6. The model also shows high performance across other benchmarks, indicating its balanced capabilities in general intelligence and real-world coding.

Additionally, GLM-5.1 can autonomously execute tasks for up to 8 hours, opening new horizons for developers of autonomous agents. It can independently navigate the full cycle from planning to testing and delivery, significantly simplifying the development process for complex systems.