Perplexity launches Portable Computer, a local AI agent for Nvidia DGX Spark
Perplexity, already known for search, advertising, a personal assistant, a browser and a cloud-based agent platform called Computer, has built a local counterpart named Portable Computer. It pairs an agent harness with an orchestrator and runs open-weight AI models directly on an Nvidia DGX Spark workstation, falling back to cloud inference only when needed.
Perplexity frames the move around cost and privacy. In its announcement post the company wrote that small open-weight models have gotten good enough for real agentic work, naming Nvidia Nemotron 3.5 Lightning (30B total parameters), Qwen 3.6 (35B) and Qwen 3.8 (27B) as examples: "These small models punch above their weight and are now capable of complex agentic workflows." For its own benchmark testing, Perplexity ran Qwen 3.8 at 27B parameters on the DGX Spark hardware.
The pitch: local inference has no per-token API fees and keeps private data on the device rather than sending it to a remote cluster. Perplexity compared Portable Computer against two other agent harnesses, Pi and Hermes, both of which are free. It says Portable Computer matched or beat them on accuracy, came out fastest on the BrowseComp and ParseBench-100 benchmarks, and used the fewest tokens across all three benchmarks tested. The company's conclusion, in its own words, is that a capable open-weight model plus local hardware and a good harness "can handle real knowledge work at near-zero inference cost without requiring sensitive data to leave the device."
The Register ties the announcement to a separate, unconfirmed report that Nvidia is contemplating a $30 billion investment in Perplexity, noting dryly that Perplexity's Portable Computer post mentions Nvidia's brand name six times. No release date, price or availability timeline was given for Portable Computer, and no individual at either company was named in the announcement.
Key facts
- Perplexity launched Portable Computer, a local AI agent that runs an agent harness and orchestrator with open-weight models on an Nvidia DGX Spark workstation, falling back to the cloud only when needed.
- Perplexity cited Nvidia Nemotron 3.5 Lightning (30B parameters), Qwen 3.6 (35B) and Qwen 3.8 (27B) as small models now capable of complex agentic workflows, and benchmarked Portable Computer using Qwen 3.8 at 27B on the DGX Spark.
- Perplexity says Portable Computer matched or exceeded rival free agent harnesses Pi and Hermes on accuracy, was fastest on the BrowseComp and ParseBench-100 benchmarks, and used the fewest tokens on all three benchmarks tested.
- The Register linked the launch to reports that Nvidia is contemplating a $30 billion investment in Perplexity, a deal still described as under discussion rather than confirmed.
- No price, release date or availability timeline was given for Portable Computer.
Why it matters
Perplexity has spent the past few years expanding from search into advertising, a personal assistant, a browser and a cloud agent platform called Computer. Portable Computer is its move into running agents locally rather than in the cloud, betting that open-weight models in the 27 to 35 billion parameter range are now good enough for real agentic work on a single workstation.
Who it affects
Anyone evaluating AI agent harnesses for cost or data-locality reasons: Perplexity positions Portable Computer against free rivals Pi and Hermes, and its pitch to potential customers is built almost entirely around benchmark comparisons with them. Nvidia is also a beneficiary either way, since Portable Computer runs on its DGX Spark hardware and Nvidia's own Nemotron model is one of the three Perplexity names as production-ready.
How to use it
Portable Computer requires an Nvidia DGX Spark workstation and runs local models such as Qwen 3.8 at 27B parameters, with cloud inference available as a fallback. Perplexity has not published a price, a release date or an availability timeline, so there is currently no way to buy or deploy it beyond what the announcement describes.
How solid is it
The benchmark claims (matching or beating Pi and Hermes on accuracy, fastest on BrowseComp and ParseBench-100, fewest tokens across all three tests) come from Perplexity's own announcement, not an independent evaluation, and the company gives no percentage or dollar figure for the cost savings it claims local inference delivers.
Risks and caveats
The reported $30 billion Nvidia investment in Perplexity is, by The Register's own account, only under discussion, not a signed deal, and carries no timeline. The Register also points out that Perplexity's announcement names Nvidia's brand six times and leans heavily on Nvidia's own Nemotron model in its examples, while running the whole system on Nvidia hardware, an alignment of interests worth keeping in mind when reading the benchmark claims. Perplexity's current privacy policy does not specifically address Portable Computer, even though it discloses that Perplexity collects substantial data through its other services.
“This local-first approach enables significant cost savings, since local inference avoids per-token API fees. It also naturally resolves the privacy and intellectual-property concerns: private tokens never need to be transmitted to remote clusters and remain safely within the boundary of the local device.”
— Perplexity, company announcement