Perplexity just shipped what a lot of enterprise AI teams have been asking for: an agent that runs entirely on hardware you own, with no per-token bill unless you explicitly escalate to the cloud. Portable Computer packages the full stack — local model, inference engine, agent harness, tool sandbox, and app connectors — as one system on NVIDIA’s DGX Spark. That’s a meaningful departure from the usual DIY approach of standing up an inference server and wiring tools by hand.
The catch is the hardware gate. You need a GB10-class box or an RTX GPU with at least 24 GB of VRAM. That puts this squarely in the territory of finance, legal, healthcare, and defense teams that already own NVIDIA workstations and have contractual reasons to keep data on-prem. General SMBs need not apply — the machine is the price of entry, and Perplexity isn’t pretending otherwise.
The design decision worth paying attention to is the escalation gate. Local-first doesn’t mean local-only. When a step needs live web access or frontier reasoning, the orchestrator stops, runs a PII classifier over the selected context, and shows you exactly what would leave the machine before asking permission. The remote model returns text guidance only — no direct access to local files, tools, or conversation history. That’s a cleaner privacy story than most hybrid setups, which tend to blur the line between local and cloud processing.
Perplexity’s own benchmark numbers tell an interesting story. On its 53-task Local Knowledge Work Bench, Computer with Qwen 3.8 27B scored 82.6%, beating the open-source Pi harness (77.6%) and Hermes (74.0%) on the identical model. PPLX 27B, Perplexity’s post-trained variant, pushed that to 85.4%. The hybrid result is more revealing: on Terminal Bench 2.1, fully local scored 59.6% at effectively zero marginal cost, while cloud escalation lifted it to 73.0% at roughly $0.415 per rollout. Claude Opus 5 alone scored 82.4% at about $0.65. Escalation narrows the gap without closing it — which feels like an honest acknowledgment of where small models still fall short. The Qwen model’s 260K-token window degrades past roughly 100K, so Perplexity engineered around that by keeping system prompts small, loading skills on demand, and exposing connectors as compact CLI tools instead of full MCP definitions. That kind of practical engineering is what separates a shipping product from a benchmark demo.
