GLM-5: The Open-Source Model for Production-Grade Coding Agents

The FriendliAI team is continuing our partnership with Z.ai, to offer Day 0 support for GLM-5. The most advanced open-source foundation model built for complex systems engineering and long-horizon agent workflows is now available on Friendli Serverless and Dedicated Endpoints.

Advances for GLM-5

GLM-5 introduces new features that strengthen its abilities to reason, execute long-horizon agentic tasks, and write frontend or backend code:

• Agentic long-horizon planning and execution: GLM-5 is purpose-built for multi-stage, long-step complex tasks. It can autonomously decompose system-level requirements with an architect-level approach, while maintaining context coherence and goal alignment across automated workflows that run for hours.
• Backend refactoring and deep debugging: GLM-5 demonstrates strong depth reasoning in backend architecture design, complex algorithm implementation, and difficult bug resolution. It includes robust self-reflection and error correction mechanisms, enabling it to analyze logs, identify root causes, and iteratively fix issues after compilation or runtime failures until the system runs end-to-end.
• Automation of Knowledge Work with Office by Z.ai: GLM-5 can turn text or source materials directly into .docx, .pdf, and .xlsx files—PRDs, lesson plans, exams, spreadsheets, financial reports, run sheets, menus, and more—delivered end-to-end as ready-to-use documents. It supports multi-turn collaboration and turning outputs into real deliverables.
• An open-source alternative with Opus-level intelligence: GLM-5 directly benchmarks against Claude Opus 4.5 in code logic density and systems engineering capability, while providing open-source deployment flexibility and strong cost efficiency.

As a result, GLM-5 delivers production-grade productivity and performance gains beyond GLM-4.7, comparable with top closed-source alternatives, like Claude Opus 4.5.

In a world with Claude Code and OpenClaw, developers desperately need an open-source model that can build complex systems and power agentic applications, not just write code. That’s why FriendliAI serves GLM-5 with the lowest latency, highest throughput and compute efficiency.

Large Model, Efficient Architecture

GLM-5 is an MIT-licensed large multimodal model pre-trained with 740 billion total parameters, 40 billion active parameters, and 28.5 trillion tokens of data. Its Mixture of Experts (MoE) architecture enables the model to retain proper context across multi-turn conversations or tasks, optimizing token utilization.

DeepSeek Sparse Attention (DSA) helps GLM-5 determine which tokens to prioritize during inference. It adapts the sparsity pattern based on input content to balance efficiency with accuracy. The team at Z.ai also incorporated slime into GLM-5, a novel asynchronous reinforcement learning infrastructure that substantially improves training throughput and efficiency for more fine-grained post-training iterations.

One-Click Deployment for GLM-5

Deploy GLM-5 with one-click on FriendliAI by creating your endpoint. Choose between:

• Friendli Dedicated Endpoints for consistent performance and guaranteed availability on reserved GPUs, or…
• Friendli Serverless Endpoints to run inference with low latency and high throughput for testing and developing AI applications

Then, all you need to do is:

1. Configure your model and compute instances
2. Enable (or disable) LoRA adapters, quantization, autoscaling, and engine configurations

API calls can be made for $1.00 per million input tokens and $3.20 per million output tokens.

Getting Started with FriendliAI

We’re currently offering up to $50,000 in credits to switch to the most performant engine. Contact our team to determine your eligibility while the GPUs are still hot! Visit this page to learn more.