Portable Computer runs Perplexity's agent stack and 27B models locally, then asks before sending harder steps to the company's cloud.
By RuntimeWire Staff · Published
Primary source: Perplexity
Why it matters #
Perplexity is turning model routing into a product moat: local hardware handles cheap, private work, while its cloud remains the paid escalation path for harder tasks.
Perplexity, led by co-founder and CEO Aravind Srinivas (@AravSrinivas), launched Portable Computer on Tuesday, moving its Computer agent stack onto NVIDIA's DGX Spark so files, models and most tool execution can stay on hardware controlled by the user.
The name oversells the hardware. Portable Computer is software for an existing desktop AI system, rather than a laptop or a Perplexity-built device. It runs the orchestrator, planner, tool router, scheduler, task queue and local search index on NVIDIA hardware, with Qwen 3.8 27B or Perplexity's post-trained PPLX 27B handling the initial workload.
Srinivas, who earned an electrical engineering degree at IIT Madras and a computer science Ph.D. at UC Berkeley, built Perplexity around a similar orchestration idea: combine models with retrieval and citations, then hide most of that machinery behind a consumer product. His IIT Madras profile says he interned at Google, DeepMind and OpenAI before joining OpenAI as a research scientist and co-founding Perplexity in 2022.
Portable Computer extends that approach from picking cloud models for an answer to deciding where each part of a longer job should run.
A product version of the value-per-watt thesis
Srinivas has been describing AI competition in increasingly physical terms. RuntimeWire reported in June that he viewed orchestration across cloud and on-device models as a way to maximize useful work from limited energy and compute.
Portable Computer turns that argument into a product. Perplexity says work completed locally does not consume its credits, making repetitive or long-running tasks less exposed to cloud usage charges. The agent can search local documents and code, act on files, continue jobs in the background and keep audio processing on the device.
That is also a distribution bet for Perplexity. The company began as an answer engine, then added browsers, agents and a Computer product capable of working across applications. Local execution gives Perplexity another place to own the orchestration layer, even when Perplexity is not supplying every model or running every inference request.
NVIDIA gets a practical workload for DGX Spark, a desktop system that began shipping in October 2025 with a GB10 Grace Blackwell processor and 128GB of unified memory. NVIDIA markets DGX Spark as a system for running large models and agents locally, giving customers a reason to buy local compute after years of AI spending flowing mainly into data centers.
Local first still has a cloud door
Portable Computer is local by default, rather than fully offline. When a task needs current web information, a connected application or stronger reasoning, the agent can ask to escalate that portion to Perplexity's cloud. Perplexity says the cloud path includes its search and research products and access to more than 15 frontier models.
The agent also connects to Google Drive, Gmail, Slack and GitHub. Those integrations make Portable Computer more useful than an isolated local chatbot, while expanding the number of systems and permissions involved in a workflow. A GitHub triage job, for example, could compare issues with bug reports from Gmail and send suggested owners to Slack.
Perplexity says users must approve transfers from the device to a cloud service. Its announcement also says code and tool execution run in isolated sandbox environments with controlled access to files and connected applications. The announcement does not explain the permission model's design, the granularity of those controls after authorization or how the agent behaves when it encounters prompt injection.
The privacy proposition therefore depends on routing and permissions as much as model location. Local inference can keep a private codebase or financial document on the machine. A connected workflow may still need to transmit selected content to another service, and the user has to understand what the agent proposes to send.
The hardware narrows the first market
The initial release is available to Perplexity Pro and Max subscribers using Linux on DGX Spark. NVIDIA says support for GeForce RTX and RTX Pro GPUs is coming, while Perplexity plans Windows support and a Nemotron 3.5 Lightning option.
That launch configuration makes Portable Computer a product for developers, technical teams and early local-AI buyers before it becomes a mainstream desktop agent. DGX Spark is compact, but it remains dedicated AI hardware with a Linux and Arm software environment. Broader RTX support would put Perplexity in front of a much larger installed base.
Perplexity has tried to remove some of the deployment friction. The company says Portable Computer can be installed through a one-click setup in its app. Its published DGX Spark model package includes the quantized Qwen checkpoint and serving configuration used for local inference.
The harness is part of the product
Portable Computer's engineering bet is that agent performance depends on the harness, tools, context management and routing policy surrounding the model. A smaller local model can complete useful work when the software limits wasted tokens and escalates only the steps that need outside compute. Perplexity has published an accompanying research post, though its precise benchmark figures could not be independently confirmed from the available material.
That hybrid design gives Srinivas a way to keep Perplexity relevant as open models improve and inference moves onto personal hardware. Customers can supply the machine and run much of the workload themselves. Perplexity keeps the interface, connectors, agent runtime and cloud escalation path.
Portable Computer is the clearest expression yet of where Srinivas wants Perplexity to sit. The company started between users and the web. It is now trying to sit between users, their files, their applications, local models and every cloud model the job may require.