Perplexity releases Portable Computer: local agents on DGX Spark, zero token cost until escalation

Perplexity is shipping Portable Computer, a local-first version of its Computer agent platform that runs the harness, scheduler, planner, tool router and coordinated models directly on an Nvidia DGX Spark workstation. The package bundles a local model, inference engine, tool sandbox and application connectors into a single system. Every task begins on the device, and steps handled by the local model carry no token cost. Only when a step requires live web access or frontier reasoning does the scheduler pause and ask for approval before sending that single step to one of 15-plus cloud models.

Hardware as the gatekeeper, models in a sealed box

The software ships as a finished product, not a preview binary, but the entry bar is rigid: a GB10 box — meaning a DGX Spark — or an RTX card with 24 GB of VRAM under the desk. At the enterprise level the target is mid-to-large companies that already own Nvidia workstations, and well-funded startups building AI at their core; small and medium businesses are effectively excluded because the machine is the price of admission. The clear verticals are finance, legal, healthcare, government and defense, and IP-heavy engineering — anywhere data residency or contractual confidentiality blocks cloud inference.

What actually runs on the device

Portable Computer is not a local chat with a file picker. Perplexity packages the local model, inference engine, agent harness, sandbox and connectors as one system, eliminating the usual work of standing up an inference server and wiring tools by hand. The user chooses between Qwen 3.8 27B and PPLX 27B, Perplexity's harness-tuned variant; Nvidia's Nemotron 3.5 Lightning, a 30-billion-parameter open-weight MoE model, is listed as "coming soon." Bringing your own model and inference server is also supported. Code and tool calls execute inside a sandbox that enforces OS-level controls on processes, file paths and network access; if the sandbox is unavailable, tool execution is disabled rather than silently falling back to an unprotected mode. Gmail, Outlook, Slack and GitHub connectors are routed through the local scheduler.

The escalation gate is the real architectural decision

Local-first is not local-only. When a step needs the live web or frontier reasoning, the scheduler pauses and asks. Before each call, the harness selects the relevant context, runs a PII classifier over it, and shows the user exactly what is about to leave the machine. The approved step is routed to a cloud model; the remote advisor returns text guidance only and never receives direct access to local files, tools or the conversation. Perplexity also engineered around small-model context limits: Qwen 3.8 27B claims a 260 thousand-token window but degrades after roughly 100 thousand, so the harness keeps a lean system prompt and toolset, loads specialized skills on demand, exposes connectors as compressed CLI tools instead of full MCP definitions, and compresses stale context mid-run.

Benchmarks: a gap that money partially closes

On the 53-task Local Knowledge Work Bench — deep research, financial analysis and document creation — which Perplexity plans to open-source, Computer with Qwen 3.8 27B on DGX Spark scored 82.6%, versus 77.6% for the open-source Pi harness and 74.0% for Hermes on the exact same model; PPLX 27B lifted the result to 85.4%. On BrowseComp the score was 66.7% versus 50.2% and 43.9%, with 51% less wall-clock time and 70% fewer tokens than Pi. On ParseBench-100 for visual document understanding: 65.1% versus 34.6% and 13.9%. The most instructive result is the hybrid one: on Terminal Bench 2.1 a full local run yielded 59.6% at near-zero marginal cost; escalation to the advisor raised that to 73.0% at roughly $0.415 per run, while Claude Opus 5 alone reached 82.4% at roughly $0.65. Escalation narrows the gap to frontier models without closing it.