Eigen RadarAI
Analysis

Portable Computer moves the agent stack on-device, not every task

Perplexity Portable Computer runs the model, planner, tool router and search index on the user's NVIDIA hardware. Local work consumes no token credits, but optional escalation to stronger cloud models and a minimum 24-gigabyte video-memory requirement define the product's privacy and cost boundaries.

Artificial Intelligence··Morning
In a sunlit home office, local work tools feed a computer while a hand controls a separate cloud connection switch.

The control plane also moves on-device

Portable Computer does more than run an open model locally. VentureBeat reports that the model, agent harness, inference engine, tools, connectors and security sandbox are packaged in one application. Computerworld says the orchestrator, planner, tool router, scheduler, durable task queue and local search index also remain on the device. Initial choices are Qwen 3.8 27B and PPLX 27B, Perplexity's post-trained version of the same base. The decision about whether a task needs the cloud therefore begins on the local control plane too.[1], [2]

Local work has a hardware floor

The product launched on Linux for Pro, Max, Enterprise Pro and Enterprise Max subscribers, with Windows support planned for September. It requires a DGX Spark or an NVIDIA RTX GPU with at least 24 gigabytes of video memory. Work completed locally consumes no cloud credits, but experts interviewed by Computerworld point to the upfront cost of a high-memory GPU, system memory and related hardware. Operating expense may fall; the product does not make the hardware free or claim to run on every computer.[1], [2]

Privacy depends on the cloud-escalation rule

Portable Computer can send a step to a cloud model for current information, browser use, connected applications or stronger reasoning. Perplexity says local execution is the default and each cloud action requires explicit approval. The company told Computerworld that content inside a local document cannot trigger escalation by itself; the user must leave local-only mode and approve the individual action. For organizations, the remaining control question is whether that approval is merely a user-interface prompt or an enforceable administrator policy.[1], [2]

References

  1. News sourceVentureBeatPerplexity puts its agent on the user's own GPU and drops the per-token bill↩1↩2↩3
  2. News sourceComputerworldPerplexity keeps its agent and GPU control plane local but cloud steps need approval↩1↩2↩3