Perplexity Portable Computer Runs a Local-First AI Agent on DGX Spark

Perplexity launched a local-first version of Computer

Perplexity launched Portable Computer on August 25, 2026, a local-first version of its Computer agent designed to run on NVIDIA DGX Spark.

The important distinction is architectural: the model, harness, conversation state, trajectory, orchestrator, planner, scheduler, tool router, durable task queue and local search index run on the user's machine by default.

When a task needs current web information, connected apps or stronger cloud reasoning, the system can escalate selectively. Perplexity says data is sent off-device only after the user authorizes that step.

The first release runs Qwen 3.8 27B or PPLX 27B locally

On DGX Spark, Portable Computer can run Qwen 3.8 27B or PPLX 27B, Perplexity's post-trained version of the Qwen model.

NVIDIA Nemotron 3.5 Lightning is listed as coming soon to the model picker.

Perplexity says the local agent can read and search files, operate across documents and code, run device actions and keep long-running jobs alive without charging per-token or per-credit inference fees for work completed locally.

The first release is available to Perplexity Pro and Max subscribers on DGX Spark and runs on Linux. Perplexity says Windows support and availability on NVIDIA RTX GPU PCs are coming later.

Local execution and cloud escalation are deliberately separated

Portable Computer does not claim that every task can stay offline.

The local orchestrator can request cloud capabilities for current web search, browser use, connected apps or one of more than 15 frontier models. App connectors include services such as Google Drive, Gmail, Slack and GitHub.

Perplexity says that when a task would send content from the device to a cloud service, the user is asked for permission first.

That boundary is important for confidential work. A local model could analyze private code, financial files or contracts without uploading the underlying data, then request cloud help only for external market information or difficult reasoning when the user approves the context that will leave the device.

The harness is optimized for smaller local models

Perplexity's accompanying research post argues that a local model should not simply be dropped into a harness designed for frontier cloud models.

The company says its local harness uses a compact core prompt, modular skills that load on demand, context compaction and compact command-line interfaces for frequently used connectors instead of exposing large MCP tool schemas to the model all at once.

Tool execution is placed inside an OS-level sandbox. Perplexity says the harness disables tool calls if the sandbox is unavailable rather than silently falling back to unsandboxed execution.

The deterministic orchestrator, not the model, retains control of the execution loop and policy enforcement. The model proposes actions; the harness decides what approved tools run and what context can leave the device.

Perplexity reports strong local-agent benchmark results

Perplexity evaluated the system with Qwen 3.8 27B on an NVIDIA DGX Spark against the Pi and Hermes harnesses.

On its internal Local Knowledge Work Bench, Perplexity reports 82.6% for Computer with the base Qwen model, compared with 77.6% for Pi and 74.0% for Hermes. Its post-trained PPLX 27B model raises the Computer score to 85.4%.

On BrowseComp, Perplexity reports 66.7% accuracy for Computer versus 50.2% for Pi and 43.9% for Hermes, while using fewer recorded tokens and less mean wall time than the compared harnesses.

On its ParseBench-100 document-understanding subset, Perplexity reports 65.1% for Computer, compared with 34.6% for Hermes and 13.9% for Pi.

These are Perplexity-run evaluations, and several comparisons depend on the company's own harness, search infrastructure, benchmark subset or internal task set. They are useful evidence about the design, but they should not be treated as independent proof of universal superiority.

Advisor escalation narrows the gap to frontier models

Perplexity also tested a hybrid setup on Terminal Bench 2.1.

The company reports that Qwen 3.8 27B in the local Computer harness scored 59.6% fully local. Allowing it to consult Claude Opus 5 as an advisor raised the score to 73.0% at an estimated API cost of $0.415 per rollout. Running Claude Opus 5 alone in the same harness reached 82.4% at $0.65 per rollout.

The pattern is notable because the cloud model does not directly gain access to local files or tools. Perplexity says the orchestrator selects context, flags sensitive information with a PII classifier, shows the user what would be sent and receives text guidance back from the advisor.

Local dictation is also designed to remain on-device

Portable Computer includes local dictation using NVIDIA Nemotron 3.5 ASR.

Perplexity says audio transcription and file actions can stay on the machine, which may be useful when working over private documents or code.

Why this matters

Agentic AI is becoming expensive in two ways: inference cost rises with long trajectories, and privacy risk rises when every document, tool call and intermediate state moves through remote infrastructure.

Portable Computer is an attempt to change that default.

Rather than forcing a choice between a fully local but weaker agent and an always-cloud frontier agent, the system uses a local-first execution boundary with explicit cloud escalation.

The main practical questions are now hardware availability, local-model quality, connector behavior, sandbox guarantees and how consistently the system asks for approval before transmitting context.

The launch is real today on DGX Spark for supported subscribers, while Windows, RTX-PC support and Nemotron 3.5 Lightning integration are upcoming rather than generally available.