Zhipu AI Releases GLM-5: 744B Open-Source Model for Agentic Workflows

Zhipu AI has officially launched GLM-5, its fifth-generation large language model, marking a significant advancement in open-source foundation models with 744 billion total parameters (40 billion active per token) and a 200,000-token context window. The model, released under the MIT License, is designed for complex systems engineering, long-horizon agentic tasks, and multi-step workflow execution, with technical specifications detailed in the company's official technical report.

GLM-5 scales from GLM-4.5's 355B parameters (32B active) to 744B (40B active), with pre-training data increasing from 23 trillion to 28.5 trillion tokens. The model integrates DeepSeek's Dynamically Sparse Attention (DSA) mechanism to maintain long-context efficiency while reducing deployment costs. Unlike previous iterations, GLM-5 incorporates an asynchronous reinforcement learning infrastructure called "slime," which improves training throughput and enables more granular post-training alignment, as noted in the technical documentation.

Performance benchmarks demonstrate GLM-5's competitive edge: it scores 50.4% on Humanity's Last Exam with tools (compared to 43.4% for Claude Opus 4.5), 77.8% on SWE-bench Verified coding tasks (approaching Claude Opus 4.5's 80.9%), and $4,432.12 in Vending Bench 2—a simulated one-year business operation—narrowing the gap with frontier models. The model also leads open-source rankings in agentic tasks, achieving 75.9% on BrowseComp with context management and 89.7% on τ²-Bench for complex reasoning.

Zhipu AI, a Tsinghua University spinoff that completed a landmark Hong Kong IPO on January 8, 2026, emphasized domestic hardware compatibility as a strategic milestone. While the company confirmed GLM-5 supports deployment on Huawei Ascend chips, it has not officially verified training hardware, though the model was developed using Huawei's MindSpore framework. This marks a critical step for China's AI infrastructure ambitions, reducing reliance on foreign chip ecosystems for large-scale model training.

GLM-5's open-source release on Hugging Face and ModelScope, coupled with MIT licensing, enables unrestricted commercial and research use. The model is also available via Z.ai's API and BigModel.cn, with compatibility for tools like Claude Code and OpenClaw. A specialized variant, GLM-5-Turbo, is set for release on March 24, 2026, targeting enhanced tool use and workflow automation at lower cost per inference call.

Industry analysts note GLM-5's significance lies in its balance of scale and efficiency. The Mixture-of-Experts (MoE) architecture—activating only 40B parameters per token—reduces computational overhead while maintaining performance. This contrasts with dense models like GPT-5.2, which typically require higher resource allocation for comparable outputs. The model's agentic capabilities, demonstrated through end-to-end coding and long-horizon planning, reflect a broader industry shift from "chat" to "work" AI paradigms.

For developers, GLM-5's 200K-token context window enables processing of entire codebases or lengthy documents in a single session, while its MIT license removes commercial barriers. However, the model's "execution-first" design—prioritizing task completion over conversational flexibility—may require adaptation for role-playing or creative writing use cases, as noted in secondary analyses.

With GLM-5, Zhipu AI has positioned itself as a key player in the open-source AI ecosystem, challenging both proprietary models and domestic competitors. The company's focus on hardware sovereignty, combined with benchmark-driven performance, underscores a strategic approach to competing in a market increasingly defined by infrastructure independence and specialized task execution.

Artūras Malašauskas is an AI Systems Integrator with 20+ years of production-grade web engineering experience. He has designed, shipped, and scaled enterprise Python/PHP systems for logistics, SaaS, and public-sector clients. For the past year, he has focused exclusively on AI integrations: deploying open-source LLMs, building generative media pipelines (image, audio, video), and engineering multi-agent workflows for real production environments. His standard: reproducibility, security, cost-efficient inference—no vaporware. He documents and evaluates emerging AI tooling, separating verified capabilities from marketing noise. Technical editor at: muza-ai.eu, ai-verslas.lt, ai-naujinos.lt Connect on LinkedIn