An agent stack that fits on a desk
Portable Computer is Perplexity's agentic system built to run entirely on local hardware instead of Perplexity's cloud. The company describes the orchestrator, planner, tool router, scheduler, durable task queue and local search index as all running on the device, alongside the agent harness, inference engine, tools, app connectors and a security sandbox. [1]
The harness itself is kept small on purpose. VentureBeat reported that Perplexity used a succinct system prompt and a small set of core tools, with additional capabilities loading on demand as skills rather than sitting permanently in context, and that the inference layer runs on vLLM with an option to point at a custom endpoint. [2]
Why the DGX Spark
The launch hardware is NVIDIA's DGX Spark, a desk-sized AI computer built around a 20-core Arm CPU and an NVIDIA GPU sharing 128 GB of unified memory. Perplexity says support for ordinary NVIDIA RTX GPU PCs is planned, and VentureBeat put the practical floor for those machines at roughly 24 GB of VRAM, about a GeForce RTX 3090 or newer. [1][2]
That memory number is the point of choosing the Spark over a laptop or a typical desktop GPU. An agent harness that keeps an orchestrator, a task queue and a 27 billion parameter model resident at once needs headroom a consumer card usually reserves for one job at a time, which is why the roomier unified-memory box is the reference machine and the RTX path is the fallback. [1]
A 27B model, stock or house-trained
Users pick between Qwen 3.8 27B, an open-weight model from Alibaba's Qwen line sized to fit local hardware rather than a frontier cloud model's hundreds of billions of parameters, or PPLX 27B, Perplexity's own post-trained version of that same base model. VentureBeat reported the post-trained version scoring 85.4 percent on Perplexity's benchmark against 82.6 percent for the untouched base model. [2]
Perplexity says PPLX 27B was post-trained specifically to finish as much of a task as it can on-device and to escalate to the cloud only when the task actually needs it, which makes the fine-tune a decision-making tweak on top of the base model rather than a straightforward capability upgrade. [1]
Asking before anything leaves the machine
When a step needs current information, browser use, a connected app or one of more than fifteen frontier models for harder reasoning, Portable Computer asks the user for permission before that content leaves the device. VentureBeat reported that the harness runs a PII classifier over the outgoing context first and shows the user exactly what would be sent. [1][2]
That gate is a data-residency feature as much as a cost control. An agent that plans and executes most of a workflow locally, then surfaces the one step that needs a frontier model with a specific description of what it would send, gives an enterprise buyer a concrete answer to what leaves the building instead of a policy promise. [2]
Zero credits for local steps, a cheaper escalation for the rest
Perplexity's pitch is that work the local model finishes has no per-credit charge at all, in contrast to the usual meter running on every tool call and reasoning step in a cloud agent loop. VentureBeat reported that on coding benchmarks, letting the system escalate to the cloud when it got stuck recovered roughly three-fifths of the gap to frontier performance for about two-thirds of the cost of running everything in the cloud. [2]
Agentic workflows multiply calls: a plan, several tool invocations, a check on each result. That is exactly the shape of workload where API metering adds up fastest, so moving the routine steps to a model that already sits on the user's own hardware removes the largest and most repetitive share of the bill, while the harder minority of steps still gets a frontier model when the local one is not enough. [2]
Sources
- Introducing Portable Computer for local-first AI
Perplexity · Aug 25, 2026
- Perplexity partners with Nvidia to launch Portable Computer, a fully local AI agent with zero token costs
VentureBeat · Aug 25, 2026