The Perplexity team has released a research version of a new orchestrator model for its local agent, Perplexity Computer. The system is based on the open-source GLM 5.2 model from China's Z.ai (formerly Zhipu AI) and has undergone post-training to operate in the Computer agent environment. A key feature is its hybrid architecture: GLM 5.2 handles the primary load of executing queries, and upon reaching its limits, escalates the task to a more powerful model, such as Opus.
Efficiency on the Edge of Cutting-Edge Solutions
Perplexity claims that the fine-tuned version of GLM 5.2 demonstrates performance close to flagship models, yet costs only 0.344x the cost of using Opus. Paired with an "advisor"—a specialized tool that determines the escalation point—the system performs at the level of Opus 4.8, but with significantly lower costs. This is not a replacement, but intelligent resource allocation: a cheap base layer for routine tasks and heavy artillery only for complex scenarios.
Technical Details and Architecture
GLM 5.2 is an open-source model with approximately 744 billion parameters, released under the MIT license. This gives Perplexity complete freedom in fine-tuning and commercial use without the restrictions typical of closed APIs. The Perplexity Computer agent system orchestrates the work of over 19 models, and the adapted GLM becomes a cost-effective foundation for most tasks. A key element is the advisor tool, which analyzes query complexity and decides when to engage a more powerful model.
Geopolitical Context and Strategy
Notably, Perplexity uses a Chinese model but hosts the adapted version on its own infrastructure in the United States. This reduces dependence on external APIs and allows full control over the post-training, routing, and execution layers. Z.ai has been on the U.S. Entity List since January 2025, but Perplexity bypasses restrictions by working with the model's open weights. Previously, the company applied a similar approach with DeepSeek R1, releasing an adapted version, R1-1776, stripped of censorship. In the case of GLM 5.2, the focus is solely on economic efficiency, not political neutralization.
Expert Conclusion
This move by Perplexity is a vivid example of how open-source models are becoming the foundation for creating hybrid, cost-effective agent systems. I believe the strategy of "cheap base layer + escalation to a flagship" is the future of AI agents, especially given the high cost of top-tier models. The next step will be the adaptation of Nemotron 3 Ultra, which will further strengthen Perplexity's position in the race for autonomous agents.