Running Perplexity’s agent locally costs $4,800
Perplexity and Nvidia have moved Computer, Perplexity’s agentic assistant, out of the cloud and onto the user’s own machine. The new Portable Computer plans and carries out tasks with a 27-billion-parameter model running on local hardware. Pro and Max subscribers got access on August 25, per the announcement on Threads.
Two models are available at launch: Qwen3.8-27B and PPLX 27B, Perplexity’s post-trained version of it. Nvidia’s Nemotron 3.5 Lightning is coming later, and users can also bring their own models and inference servers. Enterprise Pro and Enterprise Max subscribers get access too.
Computer started out somewhere very different. When it launched on February 25, its core reasoning ran on Anthropic’s Claude Opus 4.6, with Gemini handling deep research, Grok taking lightweight tasks and ChatGPT 5.2 covering “long-context recall and wide search,” Ars Technica reported. Each task ran in an isolated environment in Perplexity’s cloud. Shrinking that down to one local 27B model meant the team had to “revisit almost everything throughout the stack,” Nate Kupp, Perplexity’s vice president of Computer Enterprise and Infrastructure, told The New Stack. He said most of the engineering went into the harness, the code around the model that feeds it context and runs its tool calls. The smaller model still has to plan a task, call tools, manage files and work through multi-step jobs.
The loop is run by a deterministic orchestrator, plain code and not another model. The local model proposes an action, and the orchestrator assembles the context, enforces policy and executes approved tool calls inside an OS-level sandbox that restricts processes, filesystem paths and network access. When the sandbox isn’t available, the harness switches itself off before any tool call rather than running it unprotected, the company says.
The harness is also built around how much context the model can actually use. Perplexity says Qwen3.8-27B supports a 260,000-token window but starts to struggle beyond 100,000 tokens. So Portable Computer keeps its core prompt and tool set small and loads extra skills only when a task needs them. Frequently used connectors were turned into command-line tools, so the model doesn’t have to read their larger MCP definitions.
On the same Qwen3.8-27B base model and DGX Spark hardware, Computer scored 82.6% on Perplexity’s internal 53-task Local Knowledge Work Bench, against 77.6% for the Pi agent and 74% for Hermes, The New Stack reported. On ParseBench-100, a 100-task subset covering charts, layouts, tables, text and formatting, Computer scored 65.1%, Hermes 34.6% and Pi 13.9%. Perplexity says it plans to open-source the internal benchmark.
Every task starts on the local device. If the local model gets stuck on a step, it can ask a cloud model for advice. Before that call goes out, the harness picks the relevant context, flags anything that may be sensitive, shows the user what would leave the machine and asks for approval. The cloud model sends back text guidance only and gets no access to the device’s files or tools, and the local orchestrator works that advice into the same run. Tasks that finish locally don’t use Perplexity’s usage-based credits. Credits are charged only when the system calls the company’s cloud models.
Portable Computer ships with connectors for Google Drive, Gmail, Slack and GitHub. Web searches and connector calls leave the device, while model inference and private-document processing stay local, according to Perplexity. In one demo, the agent went through a folder of tax documents on a DGX Spark. In another, it analyzed a local CSV file and posted its findings to Slack.
Running it takes specific hardware. The easier path is Nvidia’s DGX Spark running DGX OS, which costs $4,800. The other is a PC on Ubuntu, ARM or x64, with an Nvidia RTX card that has at least 24GB of VRAM. The New Stack puts even an older RTX 3090 with 24GB at well over $1,500. Macs aren’t supported yet, and Windows support is scheduled for September.
Despite the similar name, this is a separate product from Personal Computer, the always-on agent on a dedicated Mac mini that Perplexity announced in March. Portable Computer runs the model and harness locally, but “we’re not doing computer use at this point,” Kupp said. The pitch is convenience and privacy. Developers can already put together a local model server, tools and an execution environment themselves, but with Portable Computer, Kupp said, “You don’t have to fiddle around with inference.”
The launch also comes as Nvidia is reportedly in talks to invest in Perplexity at a $30 billion valuation. Kupp said Perplexity is “very focused on Nvidia across DGX and RTX” for now, though the company is considering other hardware.
Up next
Shopify’s CEO floats banning Claude Code over AGENTS.md
Tobi Lutke says he’s weighing a ban on Claude Code at Shopify until it reads AGENTS.md and .agents/skills, because sticking to CLAUDE.md alone leaves teams with what he calls split brain problems.
Related stories
Comments
No comments yet. Be the first.
Join the conversation
Sign in to leave a comment. It appears after moderation.

