Zhipu AI Unveils GLM-4.5: Open-Source AI with Integrated Reasoning
Zhipu AI Unveils GLM-4.5: A New Benchmark in Open-Source AI
Beijing, July 28, 2025 — Zhipu AI has officially launched GLM-4.5, its next-generation flagship model designed specifically for agent applications. The model is now open-sourced on Hugging Face and ModelScope, with weights licensed under the MIT License, making it accessible to developers worldwide.
Open-Source SOTA Performance
GLM-4.5 sets a new standard for state-of-the-art (SOTA) performance in reasoning, coding, and agent capabilities. In real-world evaluations, it outperformed all other domestic models. Comprehensive testing across 12 benchmark datasets—including MMLU Pro, AIME24, and MATH500—placed GLM-4.5 third globally, first domestically, and first among open-source models.
Native Integration of Core Capabilities
The model is the first to achieve native integration of reasoning, coding, and agent functionalities. This breakthrough aligns with Zhipu AI's vision for AGI (Artificial General Intelligence), ensuring that new capabilities do not compromise existing ones.
High-Parameter-Efficiency MoE Architecture
GLM-4.5 employs a Mixture of Experts (MoE) architecture, available in two versions:
- GLM-4.5: 355 billion total parameters (32 billion activated).
- GLM-4.5-Air: 106 billion total parameters (12 billion activated).
Despite having fewer parameters than competitors like DeepSeek-R1 and Kimi-K2, GLM-4.5 delivers superior performance on multiple benchmarks.
Two-Mode Design and Training Process
The model supports:
- Thinking Mode: For complex reasoning and tool usage.
- Non-Thinking Mode: For instant responses.
The training process involved:
- Pre-training on 15 trillion tokens of general data.
- Targeted training on 8 trillion tokens for code, reasoning, and agents.
- Reinforcement learning to enhance capabilities.
Cost-Effective API Pricing
GLM-4.5 offers industry-leading pricing:
- Input: 0.8 yuan per million tokens.
- Output: 2 yuan per million tokens. The high-speed version generates up to 100 tokens per second, supporting low-latency applications.
Full-Stack Development & Ecosystem Compatibility
The model excels in full-stack tasks, from front-end website writing to back-end database management. It is optimized for frameworks like Claude Code, Cline, and Roo Code, offering one-click compatibility with Claude Code.
Multi-Platform Access & Transparency
Users can access GLM-4.5 via:
- The BigModel.cn open platform.
- Free features on Zhipu Qingyan (chatglm.cn) and z.ai. To ensure transparency, Zhipu AI has published 52 questions and Agent trajectories for verification.
The release of GLM-4.5 is poised to drive large-scale AGI adoption across industries.
Key Points:
- GLM-4.5 is the first open-source model with native reasoning, coding, and agent integration.
- Achieves SOTA performance with fewer parameters than competitors.
- Offers cost-efficient API pricing at 0.8 yuan per million input tokens.
- Supports full-stack development and major code agent frameworks.
