GLM 5.1

GLM 5.1 is a large-scale language model purpose-built for coding and long-running software engineering tasks, designed to operate as a self-directed agent rather than a quick turn-taking assistant. It uses a mixture-of-experts design with 754B total parameters and roughly 40B activated per pass, paired with DeepSeek Sparse Attention and a 200K context window that can stretch to 131K output tokens. That combination is meant to let the model keep a full engineering problem in working memory while it plans, edits, tests, and iterates, sustaining coherent behavior across frontend, backend, and systems work without constant human steering. The model also exposes a thinking mode, tool calling, and structured JSON output so it can plug into coding-agent frameworks as a drop-in reasoning engine.

GLM 5.1 sits on the GLM-5 base architecture but layers in a refined reinforcement-learning post-training pass that Z.ai highlights as the source of a 28% coding gain over its predecessor. The release emphasizes long-horizon execution, with Z.ai claiming the model can work independently on a single task for more than eight hours, planning and improving its own output along the way. On public evaluations it posts 77.8% on SWE-Bench Verified, 50.4% on HLE with tools, 92.7% on AIME 2025, 86.0% on GPQA Diamond, and scores 45.3 on Z.ai's internal coding suite, alongside a #1 placement on Vending Bench 2 among open-source models. Z.ai distributes the weights openly under an MIT-style license and frames GLM 5.1 as a flagship open alternative to closed frontier coding models, well suited for agentic engineering pipelines, research workflows, and production code generation where sustained autonomy and tool use matter more than single-turn chat.