GLM-5

GLM-5 is a next-generation foundation model engineered around a Mixture-of-Experts architecture that activates 40 billion parameters from a total pool of 744 billion, a substantial scale-up from its predecessor. The model integrates DeepSeek Sparse Attention to keep long-context processing efficient while cutting deployment costs. Designed specifically for complex systems engineering and autonomous agent workflows, GLM-5 brings architect-level reasoning to multi-stage, long-horizon tasks, maintaining context coherence across workflows that can run for hours. Its capabilities extend beyond simple code generation into deep debugging, self-correction, and full-system construction.

The model benefits from an upgraded pre-training dataset of 28.5 trillion tokens and a novel asynchronous reinforcement learning infrastructure called slime, which decouples generation from training to dramatically improve post-training efficiency. This asynchronous RL approach enables GLM-5 to learn from complex, extended interactions with hundreds of reasoning rounds and thousands of tool calls. The results place GLM-5 at the top of open-source benchmarks for coding and agentic tasks, achieving 77.8% on SWE-Bench Verified, and positioning it alongside leading closed-source models in systems engineering capability. It is purpose-built for expert developers who need autonomous execution and iterative refinement on large-scale programming challenges.