Z.ai, a leading Chinese AI company formerly known as Zhipu, has released its next-generation base model GLM-4.5, achieving SOTA performance in open-source models. Built on a fully self-developed architecture, GLM-4.5 is China’s most advanced open-source Mixture of Experts (MoE) model, setting new standards for AI performance and accessibility while also reducing costs for developers.
GLM-4.5 represents a technological breakthrough as Z.ai’s first attempts into the open-source community with a MoE architecture, available in two powerful models. The flagship GLM-4.5 boasts 355 billion total parameters, while the more streamlined GLM-4.5-Air operates with 106 billion total parameters. The GLM-4.5 series is distinguished by its comprehensive capabilities, natively integrating reasoning, coding, and agentic abilities within a single model to meet the complex demands of agentic applications.
Its performance in reasoning, coding, and agentic capabilities has been evaluated on 12 representative benchmarks. Based on the average score across all these benchmarks, GLM-4.5 has secured third place globally and the first place among both domestic and open-source models.
GLM-4.5 demonstrates remarkable parameter efficiency. Notably, GLM-4.5-Air delivers exceptional performance among 100B parameter-scale models, establishing itself as a leading model in its parameter category.
The GLM-4.5 series also demonstrates significant advantages in generation speed and pricing. API calls are priced as low as USD 0.11 per million input tokens and USD 0.28 per million output tokens. The high-speed version achieves a generation rate exceeding 100 tokens/sec.
GLM-4.5 is Z.ai’s first foundation model with native Agent capabilities built directly into its core architecture. Its “Agent-native” design makes reasoning and action intrinsic capabilities, enabling it to autonomously plan multi-step tasks, generate complex data visualizations, and manage end-to-end workflows. Moreover, integrating multiple capabilities into a single model, makes GLM-4.5 well-suited for complex agent applications, combining reasoning, perception and action within a unified framework.
Unlike closed alternatives, GLM-4.5 will be released under an open, auditable license, with options for on-premise deployment and fine-tuning, providing enterprise users with more of the control and transparency that they seek. GLM-4.5 is a truly open source alternative in an industry increasingly defined by closed, proprietary systems.
“We are setting a new benchmark with GLM-4.5, demonstrating that cutting-edge performance can be open, efficient, and affordable,” said Zhang Peng, CEO of Z.ai. “The first-principles approach to measuring AGI is to integrate more general intelligent capabilities without losing existing ones. GLM-4.5 is our first complete realization of this concept.”
A Rising Power in Open Foundation Models
Widely recognized as one of China’s most globally competitive AI companies, Z.ai has been named by OpenAI as one of the few global companies capable of building rival models at competitive prices.
In 2025, the company was included as an organization developing “notable AI models” for the first time in Stanford University’s Artificial Intelligence Index Report 2025. As a responsible actor on the global stage, Z.ai was the first among its Chinese peers to sign the Frontier AI Safety Commitments.
Since 2020, its open-sourced flagship models have accumulated over 40 million downloads worldwide. Its portfolio includes:
- GLM Series – A family of open-source language models from 355B down to 6B parameters, supporting Chinese, English, and multilingual use.
- AutoGLM – AutoGLM is China’s earliest agent model, which has surpassed the latest models from OpenAI and Anthropic in agent-related evaluations.
- Agent Ecosystem – Tools and APIs that allow organisations to build their own lightweight AI assistants, trained on proprietary knowledge and workflows.
- Flash Models – Small-footprint models that can outperform larger rivals on key benchmarks, enabling faster, cheaper, and safer deployments.
- Model-as-a-Service – An innovative MaaS platform for developers and enterprises.