Australian companies are increasingly moving beyond simple information gathering with artificial intelligence, now focusing on deploying AI agents capable of executing multi-step tasks autonomously.

Within this trend, the affordability and open nature of Chinese AI models are attracting growing attention and finding wider application in the local market.

Relevance AI, a Sydney-based startup that helps enterprises like Canva, KPMG, and Autodesk build AI agents, previously relied heavily on closed-source models from US developers such as OpenAI. These models, controlled by their developers, come with high usage costs.

In contrast, open-weight models from Chinese developers are downloadable. While their self-hosted or on-premises deployment still requires investment in computing power, servers, and maintenance, companies are not charged for each API call. Relevance AI is now shifting more tasks to open-weight models from Chinese firms like Zhipu AI and DeepSeek.

Open-weight models can be deployed on a company's own servers, in private cloud services, or other controlled environments, which helps reduce high-frequency usage costs and gives users greater control over their data and technology. The proportion of traffic handled by open-weight models on Relevance AI's platform has risen from about 5% to 7.5% at the start of the year to 20% to 25% currently.

Deepak John Joseph, an executive at multinational tech firm Pupa Clic in Australia, noted that deploying AI agents for autonomous multi-step tasks leads to a massive increase in token input and output, significantly driving up model usage fees. Compared to cloud hosting, self-hosting or on-premises solutions can substantially reduce costs. He reported a notable increase in Australian client inquiries about Chinese AI models in recent months.

Traffic data tracked by OpenRouter, a global aggregator platform for major AI models, shows that in the final week of June, the proportion of tokens processed by models from Chinese developers on the platform rose from 20% a year earlier to 48%.

Using open-weight models also keeps sensitive data within a company's own systems and prevents over-reliance on a few suppliers. It allows more researchers and security teams to inspect powerful models and identify vulnerabilities. Recently, an OpenAI model went rogue during internal testing, breaching the systems of US company Hugging Face. During its investigation, Hugging Face's team, hampered by security restrictions in some closed-source models that complicated forensic work, turned to a self-deployed open-weight model, GLM-5.2, developed by China's Zhipu AI, to conduct their analysis. This incident has sparked new industry discussions about the relationship between open-weight models, model security restrictions, and cyber defense capabilities.

The debate around AI openness is intensifying. The Australian Financial Review has argued that the focus of the AI "race" might not only be on building the smartest models, but also on deciding whether powerful models remain under the closed control of a few companies or become widely accessible. In this context, Chinese open-weight models are gaining more recognition and attention.

Dave Ramphous, CEO of Australian AI company Major Coding, believes that Chinese peers have consistently invested in model development and have expanded model accessibility through open releases. Matt Vitale, executive director of Australian AI platform New Dialogue, stated that open weights, synthetic training data, declining inference costs, and rapid model iteration are collectively driving the progress of artificial intelligence.