TL;DR — Key Takeaways
- Perplexity’s Portable Computer brings its agentic AI workflows to local hardware, allowing users to work with files and data without sending everything to the cloud.
- The system can still use Perplexity Search and cloud-based models when more advanced reasoning or current information is needed, with user permission before data is transmitted.
- Local processing could reduce AI costs because on-device workloads do not consume Perplexity Computer credits.
Perplexity has announced Portable Computer, a version of its agentic AI software that can run on a local machine.
The AI startup launched Perplexity Computer in February as a cloud-based agent that can coordinate multiple models and tools to complete multistep tasks. Portable Computer moves the models and orchestration software behind those workflows from Perplexity’s cloud to local hardware.
Portable Computer’s on-device stack includes an orchestrator, planner, tool router, scheduler, task queue and search index, Perplexity said. It can read and search local files, analyze data, work across documents and carry out longer-running tasks while keeping data off the cloud by default.
Users can also opt to tap into cloud services when needed. If a task calls for current information or more advanced reasoning, the agent can access Perplexity Search or cloud models. Perplexity said Portable Computer asks for permission before sending content from the device to a cloud service, so it doesn’t automatically share private files or other sensitive information. The system also runs code and tools in isolated sandboxes that limit access to local files and connected applications. Those safeguards could make local agents more practical for working with data like private codebases and internal documents that companies may be reluctant to send to the cloud.
Running more workloads locally could also bring down costs for some users. Perplexity’s research found that agent workflows can drive up token spending because models repeatedly reason and use tools within a single task. Perplexity said on-device work does not consume Computer credits, allowing cloud spending to be reserved for more complex work. In one coding benchmark, a local model that called Claude Opus 5 only for more difficult parts of its task cost an estimated $0.415 per task, compared with $0.65 for running Claude Opus 5 throughout.
At its launch, Portable Computer is available to Perplexity Pro and Max subscribers on Nvidia’s DGX Spark desktop AI system, with support for RTX GPU PCs coming later. Nvidia is an existing Perplexity investor and is reportedly discussing an investment in a new funding round that would value the startup at more than $30 billion.
Portable Computer currently runs Qwen 3.8 27B or PPLX 27B, a version Perplexity further trained for use in its agent system. Nvidia’s 30B Nemotron 3.5 Lightning model is expected to become another local model option later. For those who need to pick up a DGX Spark, Nvidia currently lists the 4 TB model at $4,699, with 128 GB of unified memory on its Grace Blackwell GB10 platform.
Perplexity’s research for Portable Computer also shows why cloud models are likely to be preferred for the most demanding AI workloads, at least until local models can offer comparable performance. On Terminal Bench 2.1, a benchmark for terminal-based agent tasks, the local Qwen model successfully completed 59.6% of the tasks, compared with 82.4% for Claude Opus 5. Allowing the local model to consult Opus as an advisor raised its score to 73%. For now, local AI agents like Portable Computer are unlikely to replace cloud models anytime soon, but for the work they can handle, costly cloud calls may no longer be necessary.
