GLM-5

Z.ai launched GLM-5 on 2026-02-12, scaling from GLM-4.5's 355B parameters (32B active, 23T tokens) to 744B parameters (40B active, 28.5T pre-training tokens) and integrating DeepSeek Sparse Attention to cut long-context deployment costs. The new slime asynchronous RL infrastructure enables more fine-grained post-training iterations at scale. GLM-5 targets complex systems engineering and long-horizon agentic tasks, with Z.ai reporting best-in-class open-source performance on reasoning, coding, and agentic benchmarks and narrowing the gap to Claude Opus 4.5 on CC-Bench-V2 and Vending Bench 2. Weights were released under MIT on Hugging Face and ModelScope, with API access via api.z.ai and BigModel.cn and Claude Code/OpenClaw compatibility.

Independent analyst Maxime Labonne corroborates Z.ai's February 11, 2026 GLM-5 release as a 744B-parameter MoE with 40B active parameters and a 200K-token context window enabled by DeepSeek Sparse Attention. GLM-5 reportedly tops Artificial Analysis among open-weight models and reaches first among open models on LMArena Text Arena with a score of 1452, while posting 77.8% on SWE-bench Verified, 92.7% on AIME 2026, and 86.0% on GPQA-Diamond. Labonne reports GLM-5 was trained entirely on Huawei Ascend chips using the MindSpore framework with no NVIDIA dependency, a notable constraint given Zhipu's January 2025 placement on the U.S. Entity List. Weights are MIT-licensed on Hugging Face and distributed via Z.ai API and OpenRouter, with the model beating Gemini 3 Pro and GPT-5.2 on SWE-bench Verified while trailing Claude Opus 4.5.

Z.AI's official release-notes index documents the evolution of the GLM-5 line beyond the base model, confirming successive variants released by Z.AI. The page lists GLM-5.1 (April 7, 2026), GLM-5.2 (June 16, 2026), GLM-5.3 (August 18, 2026), and GLM-5.3-Flash (August 26, 2026), establishing GLM-5 as the foundation later iterations extended. These entries also provide context about the family direction: GLM-5.3 claims a 50% gain over GLM-5.2 on Z.ai Code Bench, and GLM-5.3-Flash adds native visual capabilities across code, browsers, and GUIs. The notes serve as first-party lineage evidence for the GLM-5 model family without making claims about Amazon Bedrock availability.

InferenceX provides a third-party technical overview confirming GLM-5's core specifications: 744B parameters with 40B active, 28.5T tokens of pre-training data, integration of DeepSeek Sparse Attention (DSA) to cut training and inference cost while maintaining long-context fidelity, and the "slime" asynchronous RL infr The overview also documents the GLM-5.1 follow-up point release (repository created 2026-04-03, public release April 7, 2026), described by Z.ai as the next-generation flagship for agentic engineering with state-of-the-art performance on SWE-Bench Pro and leading improvements on NL2Repo and Terminal-Bench 2.0. The arXi