GLM-5 – Open-weights model for long-horizon agentic engineering

Summary: GLM-5 is a 744B parameter Mixture of Experts model with 40B active parameters, designed for complex systems and agentic tasks. It uses DeepSeek Sparse Attention and a new asynchronous reinforcement learning infrastructure called "slime" to optimize performance and cost. It ranks #1 among open-source models on Vending Bench 2, approaching Claude Opus 4.5's results.

What it does

GLM-5 processes long-context inputs efficiently using Sparse Attention and executes agentic tasks via an asynchronous RL framework, enabling task execution beyond chat interactions.

Who it's for

It targets developers and researchers working on complex, long-horizon agentic systems requiring scalable open-source models.

Why it matters

It narrows the performance gap between open-source and proprietary models in long-term agentic simulations while reducing computational costs.