Z.ai, the AI startup formerly called Zhipu, has unveiled its new GLM-4.5 artificial intelligence model, catching the industry’s attention with its promise of lower operating costs and greater efficiency. The announcement came on Monday, with CEO Zhang Peng sharing key details of the development during a conversation with CNBC. The company’s latest offering is making some waves in the swelling world of open source AI development.
The new GLM-4.5 model enters an already competitive segment, but Z.ai aims to shift the economics of generative AI. One of the standout aspects is its cost to operate. Zhang Peng says that, compared to the widely cited DeepSeek model, GLM-4.5 is less expensive for companies to integrate. Notably, the new model is about half the size of DeepSeek, yet it delivers similar capabilities and remains open source. Any user can download it free of charge, a move that fits with the current push among AI researchers and developers for more accessible, adaptable language models.
Zhang also pointed out a key advantage for organizations focused on efficient deployment. GLM-4.5 only requires eight Nvidia H20 chips to run, which is a marked reduction from what is generally necessary for models of this caliber. In a market where hardware availability and cost are hot topics, the potential to cut infrastructure needs in half is significant. It speaks to Z.ai’s efforts to address not just the computational power behind AI but also the costs that accompany scaling up model deployments.
There is also a symbolic element to Z.ai’s approach. The company’s shift from the name Zhipu to Z.ai reflects a broader ambition to carve out a distinct identity as the artificial intelligence industry evolves rapidly. That decision has coincided with GLM-4.5’s release, underscoring a strategy focused on collaboration and transparency through open sourcing.
The open source nature of the new model is itself a notable trend among AI startups, especially in China. Firms are betting that putting sophisticated models in the hands of more users and developers accelerates improvements while lowering barriers to adoption. But while cost and accessibility are front and center, the technical details also matter. GLM-4.5’s reduced size comes without, according to Z.ai, a hit to performance which raises questions about how future AI workloads might evolve as hardware and talent supply chains remain tight.
On the technical side, Nvidia’s H20 chips are purpose-built for AI inference, and their adoption is a telling signal about which vendors and architectures are becoming the standards for next-generation deployments. Companies like Z.ai and DeepSeek, by making their models both available for free and easier to run, are encouraging experiments and spurring innovation beyond what would have been possible in a closed-source, resource-heavy era.
Of course, the promise remains to be tested at scale, and developers will be watching closely for real-world feedback as enterprises and researchers begin to experiment with GLM-4.5. The path from lab to production can be winding, especially when it involves convincing stakeholders to switch models or architectures. If Z.ai’s efficiency claims hold true, however, it could set a new bar for cost-effective deployment, particularly for enterprises that have so far held off on committing substantial resources to generative AI because of steep upfront requirements1.
Overall, Z.ai’s latest move signals how quickly the AI field is maturing, with a shift toward not just bigger and more powerful models, but smarter, more accessible ones as well. That emphasis on reaching a broader swath of developers and businesses, while also tackling the financial realities of AI adoption, should help set the tone for the next year in AI competition.
