GLM-5

GLM-5 is positioned as an open-weight model aimed at complex systems engineering and long-horizon agentic work, reflecting a continuation of the GLM family's push toward larger-scale, agent-oriented reasoning. Compared with its predecessor, it scales from 355B parameters (32B active) to 744B parameters (40B active) and expands pre-training from 23T to 28.5T tokens, giving the model a substantially larger knowledge base to draw on. To keep that scale tractable at inference time, GLM-5 integrates DeepSeek Sparse Attention (DSA), which preserves long-context capacity while materially reducing deployment cost, making the model more practical for production agents and sustained coding workflows.

On the post-training side, GLM-5 relies on a novel asynchronous reinforcement learning infrastructure called slime, designed to raise training throughput and enable finer-grained post-training iterations. That investment shows up in the results: the model delivers notable gains over the prior generation across academic benchmarks and is described as best-in-class among open-source systems on reasoning, coding, and agentic tasks, narrowing the gap with leading frontier models. The combination of sparse attention, expanded pre-training, and an efficient RL pipeline makes GLM-5 a strong fit for teams that want a self-hostable foundation for sophisticated multi-step engineering and agent applications, while its later evolution into a flagship variant with a solid 1M-token context signals continued progress toward longer, messier coding-agent trajectories.