GLM-5: 744B LLM with Sparse Attention, Tool Use, and Long‑Context Capabilities

GLM-5, released by the ZAI organization, is a massive multilingual language model targeting complex systems engineering and long‑horizon agentic tasks. It scales up to 744 B parameters (with 40 B active) and is pretrained on 28.5 T tokens in both English and Chinese. The model integrates DeepSeek Sparse Attention (DSA) to keep deployment costs low while preserving a very long context window, and it supports tool‑calling parsers (glm47 for tool calls and glm45 for reasoning) out of the box. The README highlights its suitability for text‑generation and conversational applications, and the model is licensed under MIT.

The authors emphasize a novel asynchronous reinforcement‑learning infrastructure called *slime*, which improves fine‑grained post‑training efficiency. Benchmark tables show GLM‑5 achieving top‑tier scores on a range of academic and practical evaluations, including reasoning (HLE), coding (SWE‑bench), multilingual tasks, and tool‑augmented benchmarks (BrowseComp, Terminal‑Bench). These results position GLM‑5 among the best open‑source models for reasoning, coding, and agentic tasks.

Deployment is a first‑class use case: GLM‑5 can be served locally with vLLM, SGLang, xLLM, KTransformers, and Ascend NPU pipelines. The README provides Docker and pip installation steps, as well as detailed command‑line examples for serving the FP8 variant with speculative decoding and auto‑tool‑choice enabled. Community links (WeChat, Discord) and API access on the Z.ai platform are also provided for developers who prefer managed services.