GLM-5: Scaled Open-Source LLM for Long-Horizon Agents and Real Work

ZhipuAI introduces GLM-5, a scaled open-source LLM built for complex systems and long-horizon agent tasks, featuring 744B parameters (40B active), 28.5T tokens, and DeepSeek Sparse Attention. With the new slime asynchronous RL infrastructure, GLM-5 achieves best-in-class open-source results across reasoning, coding, and agent benchmarks, including top performance on Vending Bench 2. It is released under MIT, available via APIs and popular agents, supports document-generation workflows, and can be deployed locally across diverse hardware.

Key Points

• GLM-5 scales to 744B parameters (40B active) with 28.5T tokens and integrates DeepSeek Sparse Attention to reduce cost while maintaining long-context strength.
• A new asynchronous RL infrastructure (slime) increases training throughput, enabling more efficient and iterative post-training for higher-quality behavior.
• GLM-5 delivers best-in-class open-source performance across reasoning, coding, and agentic benchmarks, including #1 open-source on Vending Bench 2 and gains on SWE-bench and Terminal-Bench 2.0.
• The model shifts from chat to work by generating ready-to-use .docx/.pdf/.xlsx deliverables, supported by Z.ai’s Agent mode for multi-turn, tool-augmented execution.
• It is openly available under MIT on Hugging Face/ModelScope, accessible via APIs, integrates with coding agents and OpenClaw, and supports local deployment on diverse hardware.

Sentiment

No meaningful sentiment about GLM-5 was expressed. The discussion was entirely meta — focused on thread management and bot detection rather than the article's content.

In Agreement

• No substantive agreeing viewpoints were expressed on this duplicate thread

Opposed

• No substantive opposing viewpoints were expressed on this duplicate thread