Perplexity’s Computer agent can now run locally — if you can afford it
The New Stack Frederic Lardinois ● Covered by 4 sources
Perplexity moved its Computer agent onto your desktop with Portable Computer. It only runs on pricey Nvidia gear, so “local AI” still isn’t cheap.
Based on reporting by The New Stack, Frederic Lardinois — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Perplexity has brought Computer, its agentic AI assistant, down from the cloud and onto the desktop as Portable Computer, in partnership with Nvidia. The pitch is simple: local runs for privacy and convenience. The catch is just as simple: you need the right machine, and those machines are not cheap.
Right now, Perplexity is focused on Nvidia hardware. The easiest path is a DGX Spark desktop running DGX OS. A more traditional option is an Ubuntu system on ARM or x64 hardware with an Nvidia RTX card that has at least 24GB of GPU VRAM. A DGX Spark costs $4,800, and an older RTX 3090 with 24GB of VRAM is still well over $1,500. So yes, local AI is getting easier to start. No, it is not getting cheap.
Getting Computer to run locally was more than swapping in a smaller model. The agent still has to read and edit files, run shell commands, process PDFs, connect to outside services, and do it all inside a local sandbox. Nate Kupp, Perplexity’s vice president of Computer Enterprise and Infrastructure, said the team had to “revisit almost everything throughout the stack.” Most of the work, he said, went into the harness: the layer that lets the model plan tasks, call tools, manage files, and execute multi-step jobs with fewer parameters than the company’s cloud models.
Perplexity says the orchestrator that runs the agent loop is deterministic code, not another model. The local model proposes an action, while the orchestrator assembles context, enforces policy, and runs approved tool calls inside an OS-level sandbox. If that sandbox is missing, the harness shuts itself off before making any tool calls. That is a nice bit of self-defense, and a reminder that “local” still needs a lot of guardrails.
The company also says context limits shaped the design. Qwen3.8-27B can support a 260,000-token window, but Perplexity says it starts to struggle beyond 100,000 tokens, so Portable Computer keeps the core prompt and tool set small and loads extra skills only when needed. On Perplexity’s internal 53-task Local Knowledge Work Bench, the system scored 82.6% with the Qwen3.8-27B base model on DGX Spark hardware, ahead of Pi at 77.6% and Hermes at 74%. On ParseBench-100, it scored 65.1%, while Hermes got 34.6% and Pi 13.9%. Perplexity says it plans to open-source that benchmark.
When a task needs help from outside the machine, Portable Computer can ask a cloud model for advice, but the local harness decides what context to send and asks the user to approve it first. The cloud side gets text guidance, not direct access to files or tools. At launch, users can choose Qwen3.8-27B or PPLX 27B, bring their own models and inference servers, and later get Nvidia’s Nemotron 3.5 Lightning. Windows support is scheduled for September, while Mac users are out for now.
My take — AI-written commentary, not fact-checked reporting
This is the right direction, but the hardware gate is doing a lot of work here. The industry keeps calling things “local” while quietly asking for a small data center on your desk. The real story is not that agents are finally moving offline; it’s that good offline agents are becoming a premium product, as usual.
Read more about this at: The New Stack