DGX Spark runs Perplexity Computer locally, with cloud access gated by approval

Portable Computer runs models, orchestration, tools, and files on a DGX Spark, then requests approval before anything leaves for the cloud.

Reading confidential documents, analyzing a code repository, or completing a long-running task without sending every step to remote servers is the premise behind Portable Computer. This new version of Perplexity Computer moves the model, agent harness, planning, tool execution, session memory, and local file search onto the user’s own machine.

The first compatible environment is NVIDIA DGX Spark, a compact desktop computer equipped with 128GB of unified memory. Portable Computer is available to Perplexity Pro and Max subscribers, with the initial release running on Linux. NVIDIA RTX PCs and Windows support are expected to follow.

Unlike Perplexity Computer, whose execution primarily relies on remote models and sandboxes, every task begins on the DGX Spark. The local model breaks down the request, reads authorized files, calls the necessary tools, and preserves the state of the work across sessions. Repository migrations, batch document analysis, or report production can therefore run without charges tied to the number of tokens consumed by the local model.

The phrase “fully local” still needs some clarification. Portable Computer can operate without a cloud model for tasks limited to data and tools stored on the machine. Web searches, current information, Gmail access, or posting a message to Slack still require connections to external services. The product is therefore local-first rather than systematically offline.

Connectors include Google Drive, Gmail, Outlook, Slack, and GitHub. They are controlled by the local orchestrator, but the action itself necessarily reaches the infrastructure of the relevant service. Reading a document stored on the DGX Spark can remain entirely local. Sending its summary to Slack takes that message outside the machine.

The same principle applies to Perplexity Search, extended research, and cloud models. When a step exceeds the local model’s capabilities or requires external information, Portable Computer pauses the run and displays what needs to leave the device. The user can deny or approve the transfer before the task continues.

For escalation to a more capable model, the harness selects only the context considered relevant and applies a classifier that flags personal information. The user then reviews the content that would be transmitted. The remote model receives this approved context and returns text guidance only. It has no direct access to the files, tools, or full conversation stored on the DGX Spark.

This separation also keeps control of actions on the local machine. The cloud model can suggest an approach, resolve ambiguity, or help unblock a step, but it cannot manipulate local resources itself. The local model and harness decide what to do with its recommendations.

Perplexity’s terminology introduces a slight ambiguity. Its launch thread on 𝕏 refers to an “orchestrator LLM” installed locally. The technical post offers a more precise description, defining the orchestrator as deterministic code. It controls the execution loop, prepares the context, and enforces policies, while the local model proposes the next action. The interface still labels that local model as the “orchestrator.” Both components run on the machine, but they perform different roles.

The local stack also includes a planner, tool router, scheduler, durable task queue, and local search index. The index retrieves information from documents and code without sending it to a remote search system. Subagents can work simultaneously on separate parts of the same request.

Code and tools run inside isolated operating-system-level sandboxes. Processes, accessible paths, and network connections are restricted according to the policies applied to each task. The technical post states that the harness disables itself if this isolation becomes unavailable, instead of continuing with the user’s general system permissions.

At launch, Portable Computer supports Qwen 3.8 27B and PPLX 27B, a version of Qwen post-trained by Perplexity. The additional work focuses less on general knowledge than on how the model behaves inside the harness: selecting tools, reading files, verifying actions, managing context size, and deciding when external assistance is needed.

Perplexity says it trained PPLX 27B in two stages. The first uses supervised learning on the strongest synthetic task trajectories, followed by reinforcement learning intended to improve robustness. The exercises were built from representative knowledge-work situations without using real documents or personal user information.

The harness has also been simplified around the limitations of a local model. Perplexity says Qwen 3.8 27B supports a theoretical context window of 260,000 tokens but becomes less reliable beyond 100,000. The system therefore keeps its main prompt short, loads skills only when needed, and summarizes older information as a task grows.

Connectors follow the same approach. Instead of loading large MCP tool definitions into the context, the most commonly used integrations have been converted into compact command-line tools. This leaves more room for instructions, documents, and intermediate results.

Perplexity’s published evaluations attempt to measure the contribution of this architecture. On Local Knowledge Work Bench, an internal test containing 53 tasks across research, finance, document creation, engineering, compliance, and visualization, Computer scores 82.6% with Qwen 3.8 27B. The same model scores 77.6% with Pi and 74% with Hermes. PPLX 27B raises Computer’s score to 85.4%.

These figures come from Perplexity, and its internal benchmark has not yet been released. The company says the 53 tasks were held out from training and evaluated through three trials each. It plans to publish the benchmark but has not yet released the files required for independent reproduction.

On BrowseComp, Computer reaches 66.7%, compared with 50.2% for Pi and 43.9% for Hermes. The comparison does not measure the harness alone. Computer uses Perplexity’s