Summary GLM-5 addresses complex systems engineering and long-horizon agentic tasks, offering enhanced intelligence efficiency through significant scaling and novel training methodologies. Targeting researchers and engineers, it aims to close the gap with frontier models in reasoning, coding, and agentic capabilities, providing a powerful open-source alternative.

How It Works GLM-5 scales to 744B parameters (40B active) and was pre-trained on 28.5T tokens. It integrates DeepSeek Sparse Attention (DSA) to reduce deployment costs while maintaining long-context capacity. Post-training is enhanced by slime, a novel asynchronous reinforcement learning infrastructure designed to improve throughput and efficiency for fine-grained model iterations.

Quick Start & Requirements Local deployment is supported via vLLM, SGLang, and xLLM.

• vLLM: Install with pip install -U vllm --pre --index-url https://pypi.org/simple --extra-index-url https://wheels.vllm.ai/nightly and upgrade transformers. Docker image vllm/vllm-openai:nightly is also available.
• SGLang: Docker images lmsysorg/sglang:glm5-hopper (Hopper GPU) or lmsysorg/sglang:glm5-blackwell (Blackwell GPU) are provided.
• Requirements: Specific GPU architectures (e.g., Hopper, Blackwell) are implied for SGLang Docker images. vLLM requires transformers from a specific git commit.
• Links: General deployment guide, vLLM recipes, sglang cookbook.

Highlighted Details

• Achieves best-in-class performance among open-source models in reasoning, coding, and agentic tasks.
• On CC-Bench-V2, GLM-5 significantly outperforms GLM-4.7 on frontend, backend, and long-horizon tasks, narrowing the gap to Claude Opus 4.5.
• Ranks #1 among open-source models on Vending Bench 2, a long-term operational capability benchmark, completing a simulated one-year vending machine business with a final balance of $4,432.

Maintenance & Community Community channels include Wechat and Discord. API services are available on the Z.ai API Platform. A technical blog provides further details.

Licensing & Compatibility No explicit license information is provided in the README. Compatibility for commercial use or closed-source linking is undetermined.

Limitations & Caveats The technical report is stated as "coming soon," indicating potential for evolving documentation and features. Specific hardware requirements for optimal deployment are implied but not exhaustively detailed.