ALL INTELLIGENCE
智谱 AI 研究T1·Models

Zhipu AI Open-Sources GLM-5: 744B Model Rivals Claude Opus 4.5

By 智谱 AIOriginal title: GLM-5开源:从代码到工程,Agentic Engineering时代最好的开源模型
86AI Score

Zhipu AI has open-sourced GLM-5, scaling parameters from 355B to 744B (40B active) and pre-training data to 28.5T tokens. The model achieves open-source SOTA on benchmarks like SWE-bench-Verified, with coding and agent capabilities approaching Claude Opus 4.5. Key innovations include the 'Slime' asynchronous reinforcement learning framework and DeepSeek Sparse Attention, enabling efficient handling of complex engineering tasks and long-horizon agentic workflows.

Full Report

Contents8 sections

Today, we launch and open-source GLM-5. A consensus is gradually forming across academia and industry: large models are evolving from writing code and front-end components to engineering systems and completing complex tasks—shifting from "Vibe Coding" to "Agentic Engineering." GLM-5 is the product of this transformation: it achieves state-of-the-art (SOTA) performance among open-source models in coding and agent capabilities, with user experience in real-world programming scenarios approaching that of Claude Opus 4.5. It excels at complex system engineering and long-horizon agent tasks. On the globally authoritative Artificial Analysis leaderboard, GLM-5 ranks fourth worldwide and first among open-source models.

Larger Foundation, Stronger Intelligence

GLM-5’s new foundation provides a solid base for the capability evolution from "writing code" to "engineering systems":

  • Parameter Scale Expansion: Parameters have expanded from 355B (32B active) to 744B (40B active). Pre-training data increased from 23T to 28.5T tokens. The significantly larger pre-training compute has markedly improved the model's general intelligence.
  • Asynchronous Reinforcement Learning: We built a new framework called "Slime," which supports larger model scales and more complex reinforcement learning tasks, improving the efficiency of post-training RL pipelines. We also proposed an asynchronous agentic RL algorithm, enabling the model to continuously learn from long-horizon interactions and fully unlock the potential of the pre-trained model.
  • Sparse Attention Mechanism: For the first time, we integrated DeepSeek Sparse Attention. This maintains lossless performance on long-context tasks while significantly reducing deployment costs and improving token efficiency.

Coding Capabilities: Aligned with Claude Opus 4.5

GLM-5 has aligned its coding capabilities with Claude Opus 4.5, achieving SOTA scores among open-source models on widely recognized industry benchmarks. It scored 77.8 on SWE-bench Verified and 56.2 on Terminal Bench 2.0, both SOTA for open-source models, outperforming Gemini 3 Pro. In 2026, large models need to move from "being able to write" to "being able to complete," especially end-to-end completion of large-scale tasks. GLM-5 acts as a "system architect"; it is not just designed for creating beautiful demos but for stably delivering production-ready results. In our internal Claude Code evaluation suite, GLM-5 significantly surpasses GLM-4.7 (with an average improvement of over 20%) in programming development tasks such as front-end, back-end, and long-horizon tasks. It can autonomously perform agentic long-horizon planning and execution, back-end refactoring, and deep debugging with minimal human intervention, offering a user experience close to Opus 4.5.

Agent Capabilities: SOTA Long-Horizon Task Execution

GLM-5 achieves open-source SOTA in agent capabilities, ranking first among open-source models on multiple evaluation benchmarks: it delivers the best performance on BrowseComp (web browsing and information understanding), MCP-Atlas (tool calling and multi-step task execution), and τ²-Bench (planning and execution in complex multi-tool scenarios).

On Vending Bench 2, which measures a model's business management capabilities, GLM-5 achieved the top score among open-source models. Vending Bench 2 requires the model to manage a simulated vending machine business over a one-year period. GLM-5 ended with an account balance of $4,432, demonstrating business performance close to Claude Opus 4.5 and showcasing excellent long-term planning and resource management capabilities.

These capabilities are core to Agentic Engineering: the model must not only be able to write code and complete engineering tasks but also maintain goal consistency, manage resources, and handle multi-step dependencies during long-horizon tasks, becoming a truly Agentic Ready foundation model.

Domestic Chip Support for Online Inference Clusters

The GLM series of models has been loved by developers worldwide. After the global surge in demand for the GLM Coding Plan, we had to implement sales restrictions. The launch of GLM-5 relies on numerous domestic chips to ensure stable and efficient online services. Currently, GLM-5 has completed deep inference adaptation with domestic computing platforms including Huawei Ascend, Moore Threads, Cambricon, Kunlunxin, MetaX, Enflame, and Hygon. Through low-level operator optimization and hardware acceleration, GLM-5 achieves high throughput and low-latency stable operation on domestic chip clusters. We thank our domestic chip partners.

Typical Scenarios for Agentic Engineering

Scenario 1: End-to-End Application Development

After GLM-5 was launched anonymously (as "Pony") on OpenRouter, many developers used it to build applications that are truly usable, playable, and ready for deployment. Developers created side-scrolling puzzle games and other applications using GLM-5.

When GLM-5 evolves to complete the entire closed loop, programming tools must be restructured around this core capability. To that end, we are launching Z Code. Users simply need to clearly articulate their requirements; the model will automatically decompose tasks, and multiple agents will concurrently handle the full development workflow—including coding, running commands, debugging, previewing, and committing. With Z Code, users can even remotely direct desktop-side Agents via mobile devices to resolve engineering tasks that previously took a long time. Notably, Z Code itself was developed entirely with the assistance of GLM models.

Scenario 4: Direct Output of Office Documents

GLM-5 features enhanced capabilities for complex system engineering and long-horizon agentic tasks, enabling it to directly convert text or materials into .docx, .pdf, and .xlsx files. On Z.ai and ChatGLM (Zhipu Qingyan), users can have GLM-5 directly generate documents such as Product Requirement Documents (PRDs), lesson plans, exam papers, spreadsheets, financial reports, process charts, menus, and more.

Open Source and Usage Methods

Effective immediately, GLM-5 is open-sourced on both Hugging Face and ModelScope, with model weights released under the MIT License. GLM-5 has been included in the Max user plan, with support for Pro users expected within five days. We will gradually expand availability to ensure more users can experience and utilize GLM-5. The GLM Coding Plan supports mainstream development tools such as Claude Code and OpenCode. The GLM Coding Plan also upgrades the Agentic Engineering experience: Official integration with OpenClaw allows configuration in just a few simple steps to quickly initiate Agent workflows; Pro/Max users receive a limited allocation of AutoGLM-OpenClaw, which enables connecting cloud-based personal AI assistants to Feishu (Lark) for executing long-running tasks in office scenarios; new "GLM in Excel" benefits include an AI plugin natively adapted for the Excel environment, supporting natural language interaction via the sidebar to deeply empower data processing and spreadsheet workflows (during the Beta phase, only Max users can redeem this via plan credits).

  1. Official API Access BigModel Open Platform: https://docs.bigmodel.cn/cn/guide/models/text/glm-5 Z.ai: https://docs.z.ai/guides/llm/glm-5 OpenClaw Integration Documentation: https://docs.bigmodel.cn/cn/coding-plan/tool/openclaw

  2. Online Experience Z.ai: https://chat.z.ai ChatGLM App/Web Version: https://chatglm.cn

  3. Open Source Links GitHub: https://github.com/zai-org/GLM-5 Hugging Face: https://huggingface.co/zai-org/GLM-5 ModelScope: https://modelscope.cn/models/ZhipuAI/GLM-5

  4. Agent AutoGLM: https://autoglm.zhipuai.cn/?channel=AutoGLM_OpenClaw&redeem_modal_open=1 Z Code: https://zcode.z.ai/cn

  5. Blog https://www.zhipuai.cn/zh/research/153