Skip to content
Search

GLM-5

lastmod 2026-02-12created 2026-02-121 edits

A model targeting complex systems engineering and long-horizon agent tasks

About

  • A large language model with improved AI efficiency and long-horizon task performance
  • Compared with its predecessor, scaled up to 744 billion parameters (40 billion active), with pre-training data increased to 28.5 trillion tokens
  • Integrates DeepSeek Sparse Attention (DSA) to cut deployment costs while keeping long-context capability
  • Raises training efficiency through a new asynchronous reinforcement learning infrastructure, ‘slime’, and records top-tier performance on various benchmarks
  • Released as open source and available on Hugging Face, ModelScope, the Z.ai platform, and more; compatible with Claude Code and OpenClaw

GLM-5 overview

  • GLM-5 is a model designed for complex systems engineering and long-horizon agent tasks
    • Parameters increased from 355 billion (32 billion active) in GLM-4.5 to 744 billion (40 billion active)
    • Pre-training data expanded from 23 trillion to 28.5 trillion tokens
  • Integrates DeepSeek Sparse Attention (DSA) to greatly cut deployment costs while keeping long-context capability
  • Introduces an asynchronous reinforcement learning infrastructure called slime, improving training throughput and efficiency and enabling fine-grained post-training iterations

Performance improvements and benchmark results

  • GLM-5 shows overall performance gains over GLM-4.7, approaching the level of Claude Opus 4.5
  • Strong results on the internal evaluation set CC-Bench-V2 across frontend, backend, and long-horizon tasks
  • Ranked first among open-source models on Vending Bench 2, reaching a final balance of $4,432 in a one-year vending machine business simulation
  • World-class open-source performance across reasoning, coding, and agent tasks
    • e.g., SWE-bench Verified 77.8, BrowseComp 62.0, τ²-Bench 89.7
  • Narrows the gap with top models such as GPT-5.2 and Gemini 3.0 Pro

Open-source release and access

  • GLM-5 is released under the MIT license, with model weights downloadable from Hugging Face and ModelScope
  • Available as an API on Z.ai, BigModel.cn, api.z.ai, and more
  • Compatible with Claude Code and OpenClaw, so it can be integrated into various development environments
  • Free trial on the Z.ai platform

Office and document generation

  • GLM-5 aims for a shift “from chat to work”, acting as an office tool for knowledge workers and engineers
  • Converts text or source material directly into .docx, .pdf, and .xlsx formats to produce finished documents such as PRDs, exam papers, financial reports, and menus
  • The Z.ai application offers an Agent mode that supports PDF/Word/Excel generation, allowing multi-turn collaboration

Developer and deployment support

  • GLM Coding Plan subscribers get access to GLM-5 in stages
    • Max plan users can enable it immediately with the model name "GLM-5"
    • GLM-5 requests consume more quota than GLM-4.7
  • Provides the Z Code agent development environment for users who prefer a GUI
  • Through the OpenClaw framework, GLM-5 can be used as a personal assistant agent that works across apps and devices

Local deployment and hardware compatibility

  • GLM-5 supports inference frameworks such as vLLM and SGLang, with deployment instructions on the official GitHub
  • Also runs on non-NVIDIA chipsets (Huawei Ascend, Moore Threads, Cambricon, Kunlun, MetaX, Enflame, Hygon, etc.)
    • Achieves reasonable throughput through kernel optimization and model quantization

See also

Favorite site

author Zer0 sitemap