GLM-5.1

GLM-5.1 is the flagship next-generation open-source model from Z.ai (Zhipu AI), designed primarily for agentic engineering tasks and AI-assisted programming. It uses a Mixture-of-Experts (MoE) architecture with a DeepSeek Sparse Attention (DSA) mechanism — 744B parameters in total, with 40B active per token (256 routed experts + 1 shared, 8 active per token).

Compared to its predecessor (GLM-5), the model shows substantial improvements on agentic benchmarks: SWE-Bench Pro (58.4%), BrowseComp (68.0%), and CyberGym (68.7%). A key characteristic of GLM-5.1 is its ability to operate effectively over long horizons — the model independently plans, runs experiments, analyzes results, and iterates on its strategy across hundreds of rounds and thousands of tool calls. It supports thinking mode (chain-of-thought reasoning, enabled by default), function calling, and code generation.

The model is released under the MIT license, with weights available for download on Hugging Face (zai-org/GLM-5.1) and ModelScope (ZhipuAI/GLM-5.1). Local deployment requires a GPU cluster (at least 8× H800/A100-class GPUs; supported via vLLM, SGLang, KTransformers, and xLLM). It is also available as an API service through the Z.ai platform. The weights were trained on Huawei Ascend 910B hardware.