GLM-5 Technical Report: Open-Source Model Built for Agentic Engineering

Zhipu AI

Feb 18, 2026 · Updated Apr 25, 2026

The z.ai team released the GLM-5 technical report covering three training innovations that achieve state-of-the-art among open-source models on software engineering benchmarks. Dynamic sparse attention cuts training and inference costs while preserving long-context fidelity for multi-step agentic coding.

GLM-5, released by the z.ai team, is an open-source foundation model built for autonomous software engineering rather than simple code completion. Three innovations define it: Dynamic Sparse Attention (DSA) for cost-efficient long-context handling, an asynchronous RL infrastructure that decouples generation from training for faster post-training, and agent RL algorithms that teach the model to learn from complex multi-step interactions.

The result is state-of-the-art performance among open-source models on major benchmarks, with the biggest gains in real-world software engineering - end-to-end tasks requiring planning, writing, and iteration across a codebase.

For developers in the GLM ecosystem (z.ai runs a Claude Code-compatible API), GLM-5 is the next generation. The code, models, and full technical report are publicly available.

View the full update on arxiv.org

Z.ai

@Zai_orgFeb 18

Presenting the GLM-5 Technical Report! https://t.co/CGjxEISvFK After the launch of GLM-5, we’re pulling back the curtain on how it was built. Key innovations include: - DSA Adoption: Significantly reduces training and inference costs while preserving long-context fidelity - Asynchronous RL Infrastructure: Drastically improves post-training efficiency by decoupling generation from training - Agent RL Algorithms: Enables the model to learn from complex, long-horizon interactions more effectively Through these innovations, GLM-5 achieves SOTA performance among open-source models, with particularly strong results in real-world software engineering tasks.

338

View on X

Every HeadsUpAI update is written based on its original source and reviewed before it's published. Read our editorial standards →

See all AI news & updates from Zhipu AI →

Keep reading

Z.ai Launches GLM-5-Turbo, Optimized for Agent Tasks from the Training Phase

Z.ai released GLM-5-Turbo, a fast GLM-5 variant optimized for agent scenarios since the training phase. It targets what agent deployments need: reliable tool calls, complex instruction decomposition, and stable execution through persistent tasks. Available via the Z.ai API and OpenRouter.

Fireworks AI Adds GLM 5.1 Training to Build Long Horizon Coding Agents

Fireworks AIApr 28

Fireworks AI Adds GLM 5.1 Training to Build Long Horizon Coding Agents

Fireworks AI added Z.ai's GLM 5.1 to its training platform, supporting supervised fine-tuning and direct preference optimization with a 200K context window. This allows developers to customize the flagship agentic model for multi-hour autonomous tasks without the numerical drift common in fragmented training and inference stacks.

OpenCode Integrates GLM-5.1 Into Go With Zero Data Retention Privacy

OpenCodeApr 8

OpenCode Integrates GLM-5.1 Into Go With Zero Data Retention Privacy

OpenCode added Z.ai's new GLM-5.1 model to its OpenCode Go platform, featuring a zero-retention policy for user data. This allows developers to use a frontier-level model for agentic engineering without their proprietary code being stored or used for future training.