GLM-5 - Zhipu Open Source's next-generation flagship AI model, comparable to Opus 4.5.

What is GLM-5?

GLM-5 is ZAI's next-generation flagship AI model, with its parameter size expanded from 355B in GLM-4.5 to 744B (40B activations), and pre-trained data reaching 28.5T tokens. This model is the previously mysterious "Pony Alpha" model that topped the OpenRouter popularity charts. Designed specifically for complex systems engineering and long-range agent tasks, it integrates DeepSeek Sparse Attention to reduce deployment costs and utilizes a self-developed slim asynchronous RL infrastructure to improve training efficiency. In the Artificial Analysis leaderboard, GLM-5 ranks fourth globally and first among open-source models. The model supports generating Office documents, is compatible with tools like Claude Code, and supports deployment on domestic chips such as Huawei Ascend, Moore Threads, and Cambricon. The model is available for testing on the z.ai website and the BigModel.cn platform, and its API is now open.

Main functions of GLM-5

• Complex Systems EngineeringThe model can handle multi-level technical tasks such as front-end development and back-end architecture design, and supports the entire process of engineering delivery from requirements analysis to code implementation.
• Long-range agent tasksIt possesses long-term planning and resource management capabilities, enabling it to make independent decisions and achieve goals in simulated business scenarios such as Vending Bench 2, which require continuous operation for a year.
• Intelligent document generationIt supports directly converting text or raw materials into formats such as .docx, .pdf, and .xlsx, and outputting professional documents such as PRDs, financial reports, and lesson plans that can be used directly.
• Multi-tool collaborationIt is compatible with mainstream development toolchains such as Claude Code and OpenClaw, enabling automated operation and collaboration across applications.

Technical Principles of GLM-5

• Large-scale pre-training expansionThe model parameters have been expanded from 355B (32B activation) to 744B (40B activation), and the pre-training data has been increased from 23T to 28.5T, using more computing power to enhance the general-purpose intelligent base.
• Slime, an asynchronous reinforcement learning infrastructureOur self-developed asynchronous RL training framework addresses the efficiency bottleneck in reinforcement learning for large-scale language models. This infrastructure supports parallel reward calculation and policy updates, enabling finer-grained post-training iterations and effectively narrowing the gap between pre-training capabilities and actual performance.
• Sparse attention mechanism This is the first time that DeepSeek Sparse Attention has been integrated, which significantly reduces token consumption and deployment costs in agent scenarios while maintaining the effect of long texts without loss.
• Deeply adapted to domestic computing powerCompleted the underlying operator optimization and hardware acceleration of domestic chips such as Huawei Ascend, Moore Threads, Cambricon, Kunlun Chip, Pingtouge, and Muxi, achieving high throughput and low latency inference.

GLM-5 performance

• reasoning ability
  • Humanity’s Last Exam score was 30.5%, close to Kimi K2.5’s 31.5%, and significantly better than GLM-4.7.
  • AIME 2026 I reached 92.7%, on par with DeepSeek-V3.2, while HMMT Nov. 2025 led most competitors with 96.9%.
  • The score was 86.0% on the GPQA-Diamond Expert Reasoning Test and 82.5% on the IMO AnswerBench.
• Programming skills
  • The SWE-bench Verified real software engineering test achieved a score of 77.8%, and the multi-language version achieved 73.3%, both of which are about 4 percentage points higher than GLM-4.7.
  • Terminal-Bench 2.0 achieved a terminal operation benchmark of 56.2%, which was further improved to 61.1% in the Claude Code environment, far exceeding GLM-4.7.
  • The CyberGym cybersecurity test score was 43.2%, nearly double that of GLM-4.7's 23.5%, demonstrating the system's ability to attack and defend complex systems.
• Agent and Tool Usage
  • Vending Bench 2's one-year business simulation ranks first among open-source models with a balance of $4,432, close to Claude Opus 4.5.
  • BrowseComp achieves 62.0% on web browsing tasks, which can be increased to 75.9% by combining context management strategies, surpassing Kimi K2.5.
  • The multi-domain tool calls of τ²-Bench reached 89.7%, MCP-Atlas common set reached 67.8%, and Tool-Decathlon reached 38.0%.

• Overall RankingArtificial Analysis's authoritative ranking: Fourth globally, First in open source.

How to use GLM-5

• Online experienceVisit the z.ai website and manually select the GLM-5 model to try the Chat mode or Agent mode for free. The latter supports multi-tool collaboration and document generation. Access is also possible through the BigModel.cn platform or the Z.ai API service, compatible with OpenAI format interfaces.
• Local deployment
  • Download BF16/FP8 weights from HuggingFace and run them using the vLLM, SGLang, or xLLM framework, supporting 8-card parallel inference.
  • In non-NVIDIA environments, deployment can be achieved using domestically produced chips such as Huawei Ascend and Moore Threads, with official optimization solutions provided.
• Development tool integration
  • You can call it by configuring the model name as "GLM-5" in tools such as Claude Code, OpenCode, Kilo Code, and Roo Code.
  • Users who subscribe to the GLM Coding Plan can enable it directly, or remotely control multi-agent collaboration through the Z Code visualization environment.

GLM-5 project address

• Project official websitehttps://z.ai/blog/glm-5
• GitHub repositoryhttps://github.com/zai-org/GLM-5
• HuggingFace model libraryhttps://huggingface.co/zai-org/GLM-5

Application scenarios of GLM-5

• Complex Systems EngineeringSupports end-to-end delivery of large-scale projects, enabling independent completion of the entire process from requirement decomposition, architecture design, code implementation, and deployment.
• Legacy system reconstructionThe model can deeply understand the existing codebase and perform backend architecture optimization and modernization.
• Deep debugging and repairThe model can analyze logs, locate root causes, and iteratively fix stubborn bugs until the system is running stably.
• Intelligent assistantThe model can automatically perform scheduled tasks such as searching, organizing, and publishing 24/7, becoming the user's digital intern.
• Business decision optimizationDemonstrates long-term planning and resource management capabilities in simulated business scenarios, enabling intelligent strategy formulation.