glm-5.1

GLM-5.1 is built on a 744-billion-parameter Mixture-of-Experts architecture, activating 40 billion parameters per token, which lets it scale up computational power without costing that same amount for every inference. The model emerged as a post-training refinement focused on coding and agentic workflows, designed from the ground up to handle long-horizon tasks that stretch across hours rather than minutes. Where most assistants work in short bursts, GLM-5.1 can independently plan, execute, and iteratively improve on a single objective for over 8 hours, making it well suited for autonomous agents and coding assistants that need to see complex engineering tasks through to completion rather than stopping at the first answer.

The model's training emphasis on agentic engineering shows up most clearly in benchmark results—it achieved a score of 58.4 on SWE-Bench Pro, surpassing both GPT-5.4 and Claude Opus 4.6, and demonstrated coding performance that reached roughly 94.6% of Claude Opus 4.6 levels. These gains appear rooted in specialized post-training rather than architectural changes, since GLM-5.1 shares the same base structure as GLM-5. The MIT license on the model weights means developers can run it locally or build custom agents on top of it, while tool-calling capabilities and support for multiple thinking modes give it the flexibility needed for real-world development pipelines. For teams building autonomous coding agents or long-running development workflows, GLM-5.1 represents a direction where models are increasingly expected to sustain effort across extended sessions rather than simply answering one prompt at a time.