China’s Zhipu Releases Open-Source GLM-4.5 with 355B Parameters and Agentic Design

• Zhipu (Z.ai) Launches GLM-4.5 with Agent-Native Architecture

  On July 28, 2025, Z.ai (formerly Zhipu) launched GLM-4.5, its largest and most advanced model to date. GLM-4.5 features a native agent framework, embedding reasoning, perception, and action directly into the model’s architecture. This design enables autonomous multi-step execution, complex data visualization, and contextual decision-making—optimizing it for AI agents and enterprise automation tasks.

• Two Variants: GLM-4.5 and GLM-4.5-Air

  Z.ai released two configurations:

  • GLM-4.5 (flagship) with 355 billion parameters

  • GLM-4.5-Air, a lightweight version with 106 billion parameters

    This dual-track release gives enterprises flexibility between high-capacity deployments and resource-efficient alternatives.

• Performance Ranking: Global No. 3, Open-Source No. 1

  Based on results across 12 international AI benchmarks, Z.ai reports that GLM-4.5 ranked #3 globally, and #1 among all domestic and open-source models. Benchmarks included reasoning, coding, and agentic capabilities. These scores aim to demonstrate competitive parity with models from OpenAI, Anthropic, and DeepSeek.

• Token Pricing: Most Affordable Model in Its Class

  Z.ai’s pricing structure for GLM-4.5:

  • $0.11 per million input tokens

  • $0.28 per million output tokens

    In comparison:

  • DeepSeek R1: $0.14 input / $2.19 output

  • Kimi K2 (Moonshot): $0.15 input / $2.50 output

    These rates position GLM-4.5 as the lowest-cost high-performance LLM available, especially when combined with its open-source nature.

• Hardware Efficiency: Runs on Just 8 Nvidia H20 Chips

  GLM-4.5 operates with only 8 Nvidia H20 chips, designed for China in compliance with U.S. export restrictions. Z.ai emphasized that the model is half the size of DeepSeek’s system and requires fewer compute resources, offering a strategic edge in efficiency and scalability.

• Open-Source Release & Quantization Support

  GLM-4.5 is fully open source and compatible with quantization, making it deployable across a broad range of enterprise-grade and local hardware setups. This promotes accessibility for developers and enterprises seeking low-cost, high-performance AI without vendor lock-in.

• IPO Plans & Strategic Investor Backing

  Founded in 2019, Z.ai is reportedly preparing for an IPO in Greater China, according to Bloomberg. The firm has raised over $1.5 billion from Alibaba, Tencent, Qiming Venture Partners, Prosperity7 Ventures (Aramco-backed), and municipal investment funds from Hangzhou and Chengdu.

• U.S. Entity List Inclusion and OpenAI Warning

  Z.ai (Zhipu) is on the U.S. Commerce Department’s Entity List, limiting collaboration with U.S. tech companies. In June 2025, OpenAI publicly warned about the rapid progress of Chinese AI firms and specifically highlighted Zhipu, noting its expansion into government contracts across multiple Chinese provinces.

• China’s Expanding AI Ecosystem

  China has now surpassed 1,500 large language models, with Z.ai’s release seen as a key move in a broader open-source acceleration strategy. According to Neil Shah (Counterpoint Research) and Prabhu Ram (CyberMedia Research), GLM-4.5’s affordability and architectural design offer serious alternatives to U.S. models in enterprise AI adoption.

• Enterprise Impact and Strategic Use Cases

  GLM-4.5 is designed for enterprise-level agent deployment, particularly in multilingual processing, workflow automation, and regionalized services. Its compatibility with enterprise systems and open availability make it a strategic tool for countries and businesses seeking cost-effective and sovereign AI infrastructure.

• Security and Geopolitical Tensions

  Despite technical merits, Chinese LLM adoption raises concerns over data security, regulatory compliance, and geopolitical risk, especially in the U.S., EU, and India. Analysts call for transparent, multilateral evaluation frameworks to ensure secure integration of models like GLM-4.5 in critical systems.

• China’s AI Momentum at WAIC 2025

  At the World Artificial Intelligence Conference (WAIC) held on July 26, 2025, multiple Chinese firms showcased major releases:

  • Tencent: HunyuanWorld-1.0, a 3D scene generation model

  • Alibaba: Qwen3-Coder, a powerful code generation LLM

    These reflect a coordinated industry push for global competitiveness through open-source AI innovation.

Why This Matters:

Zhipu’s GLM-4.5 represents a critical convergence of cost-efficiency, scale, and technological autonomy. It reshapes enterprise AI economics by reducing hardware dependency, broadening access to advanced LLMs, and challenging the dominance of Western platforms like OpenAI and Anthropic. As China’s AI ecosystem matures rapidly through open-source innovation, global tech dynamics may shift, raising strategic questions around trust, standardization, and sovereignty in the AI stack.