American company Perplexity has taken a significant step in optimizing its artificial intelligence operations by introducing a research version of a new neural network orchestrator for its local Perplexity Computer agent. The foundation of this innovative solution is the open-source Chinese model GLM 5.2 from the company Z ai, which American engineers carefully fine-tuned specifically for an agentic environment. According to the developers, this hybrid approach allows achieving performance comparable to top-tier closed-source analogs, while reducing computational costs by almost three times. This is achieved because the adapted GLM 5.2 takes on the execution of most routine and basic tasks, acting as a cheap and efficient working layer of the entire system.

A key architectural feature of the innovation is a special advisor tool, which acts as a smart dispatcher. It continuously monitors the complexity of incoming requests and at the moment the base model reaches the limit of its capabilities, automatically hands over to more powerful and expensive neural networks. The Perplexity Computer agent itself is already capable of orchestrating the work of more than nineteen different models, ensuring seamless execution of complex user tasks through other AI systems. As the next strategic step, the company has announced the adaptation of another open-source model – Nemotron 3 Ultra, confirming Perplexity's course toward creating a flexible and economically viable infrastructure for AI agents.

The geopolitical and technological context of this decision deserves special attention, as Z ai (formerly known as Zhipu AI) has been on the US sanctions list since the beginning of last year. Nevertheless, Perplexity does not just use Chinese developments, but downloads the adapted version of GLM 5.2 to its own servers in the US, fully controlling the post-training and routing processes. This allows the company to minimize dependence on external closed APIs and avoid censorship issues, which it successfully fought last year when releasing the adapted version of DeepSeek R1-1776. In this case, the focus is shifted from political neutralization to pure economics: using open models with MIT licenses allows Perplexity to create powerful agentic systems without the restrictions and huge financial costs characteristic of working with proprietary AI giants.