Perplexity's Portable Computer Moves AI Agents off the Cloud and onto Your Desk

Perplexity is pushing its agent platform off the cloud and onto hardware sitting on your desk. The company launched Portable Computer, a local-first version of Perplexity Computer that runs the entire agent runtime, including the orchestrator model, subagents, and tool harness, on NVIDIA DGX Spark. Developed with Nvidia, it is one of the most aggressive attempts yet to move serious agent workloads off hosted APIs and onto local silicon.

Every task starts on the device. If a step needs frontier reasoning or live web data, the local orchestrator pauses and asks the user before routing that specific step to one of 15+ cloud models, after a PII check. The handoff is text-guidance only, so sensitive files never leave the box.

What actually runs on the Spark

Portable Computer ships as a packaged system rather than a bare local LLM. The agent harness, orchestrator, planner, tool router, post-trained model, inference engine, tool sandbox, and app connectors all install together. Work handled by local models carries no per-token charge.

At launch you can pick between two 27B-parameter models, with a third on the way:

• PPLX 27B, a Qwen variant that Perplexity has post-trained on its own agent harness
• Qwen 3.8 27B, running at 3-bit quantization, a 17.4 GB download that needs 32 GB of RAM
• NVIDIA Nemotron 3.5 Lightning, coming soon at 4-bit quantization, 19 GB download, 36 GB RAM

Connectors keep this from being a walled garden. Gmail, Outlook, Slack, GitHub, and Google Drive all route through the local orchestrator, and Perplexity Search can be called when the agent needs the open web. Installation on a Spark running Linux is an apt-get install perplexity away once Perplexity's package repo is added.

The hardware bill

The whole scheme leans on DGX Spark being a serious little box. Nvidia's machine is built around the Grace Blackwell GB10 platform: a 20-core Arm CPU and Blackwell GPU sharing 128 GB of unified memory, delivering 1 petaflop of AI performance in a form factor that sits on your desk rather than in a server rack. Portable Computer wants at least 1 TB of storage on top of that.

The capability comes at a price. Part of the appeal of the Mac Mini for local agent projects like OpenClaw was cost per compute: a Mac Mini with an M4 chip runs about $900. A DGX Spark will set you back roughly 5x that, though it is a one-time fee. On the software side, Perplexity Pro runs $20 per month, or roughly $17 per month on an annual plan. The Max tier costs $200 per month, dropping to around $167 per month with annual billing.

How the benchmarks shake out

Perplexity's research blog published numbers on the tradeoff between local-only runs and hybrid runs that escalate hard steps to a frontier model. The local setup is competitive, and the hybrid setup gets close to a pure cloud frontier model at a fraction of the cost.

Setup	Terminal Bench 2.1	Cost per rollout
Portable Computer, fully local	59.6%	~$0
Portable Computer with cloud adviser	73.0%	~$0.415
Opus 5 alone	82.4%	~$0.65

On Perplexity's 53-task internal benchmark, PPLX 27B scored 85.4% versus 77.6% for Pi on the same base weights, suggesting the harness-specific post-training is doing meaningful work. In a live demo, a 27B Qwen model at full GPU utilization on a DGX Spark reviewed investor documents and flagged cases of unnecessary fees while the cloud credit tally sat parked at zero.

Why local, why now

The strategic logic goes beyond privacy theater. Long-running agent workflows, repo-scale code migrations, batch document analysis, and deep research loops rack up eye-watering token bills on hosted APIs. Because work handled by local models carries no per-token charge, verification loops and large-scale migrations become economically sane on owned hardware.

That reframes a DGX Spark purchase as amortized inference cost. If you were going to spend hundreds a month on frontier API calls anyway, buying the hardware once and pushing most steps local starts to pencil out. Regulated industries like finance, legal, and healthcare also get a story where sensitive documents can be processed by an agent without leaving the machine.

The catches worth knowing

The rollout is deliberately narrow. Availability is Linux-first for Pro, Max, Enterprise Pro, and Enterprise Max subscribers. Windows follows in September, and macOS is not on the roadmap. Only one DGX Spark is supported at launch; clustering is on the roadmap but not shipped. Support for Linux boxes with NVIDIA RTX GPUs is also coming, which will lower the hardware floor considerably.

There is a real capability tradeoff too. Perplexity Computer in the cloud fans out across roughly 19 different models depending on the task. Portable Computer starts with two 27B models on device, using cloud escalation as the pressure valve. For raw reasoning quality on a hard, single-shot question, the frontier cloud path still wins, as the benchmarks confirm. What is new is being able to keep 60 to 80 percent of the work local and only pay for the hard parts.

What it unlocks

For anyone building agent workflows, the interesting thing is the shape of the stack. A local orchestrator that plans, holds session state, executes tools in a sandbox, and knows when to phone a friend is what most in-house agent frameworks are converging on. Having that harness plus a competent 27B model plus per-step approval gates as a first-party product on a well-defined piece of hardware is a template competitors will have to answer.

Near-term winners are power users with existing DGX Sparks, Perplexity Max subscribers who were feeling the credit ceiling, and enterprises that could not adopt cloud agents for compliance reasons. Near-term losers are pure-API agent products whose entire moat was billing per token for work a 27B model on a desk could handle for free.