Perplexity launched hybrid compute for its Computer agent platform today, routing sensitive data to local Apple silicon Macs while frontier cloud models handle planning, web research, and heavy reasoning. The system works on any Mac running macOS 15 or later and is available now to enterprise customers, Pro subscribers, and Max subscribers. At launch, users choose from three local models: Google Gemma E4B, Alibaba Qwen3.6 35B-A3B, or a Perplexity post-trained version of Qwen3.6 35B, which the company recommends.
The architecture's core is a company-trained PII classifier called Privacy Gate, which runs on-device and scans for names, addresses, account numbers, and secrets before any token reaches the cloud. When it flags sensitive content, the user decides whether that portion runs locally or gets shared. Tokens processed locally cost nothing against Perplexity's credit system. Only cloud orchestration and delegation draw credits. A 40-minute private equity demo ran a financial model against confidential projections, benchmarked public comparables, and produced a fifth-draft investment committee deck with no human input.
The piece is worth reading in full for two reasons. First, the Qwen question is unresolved: Perplexity's recommended local model is Chinese-developed, and Staff's argument that open-weight local inference eliminates geopolitical risk deserves scrutiny from any enterprise security team. Second, the training data question is explicitly unanswered: a company spokesperson said Perplexity is not using data for post-training globally but promised to follow up on non-enterprise accounts. Those are not footnotes. They are the story.
[READ ORIGINAL →]