Perplexity Portable Computer: Full Agent Harness on DGX Spark, $0 Local Steps
In this article
Perplexity has shipped Portable Computer, a production release of its agentic Computer platform that runs the full agent harness — orchestrator, planner, tool router, and post-trained models — directly on NVIDIA DGX Spark hardware. The inference engine, tool sandbox, and app connectors ship as one packaged system; every task begins on the device. Work handled by local models carries no per-token charge, which restructures the economics of long-running agentic loops for teams operating on owned hardware.
For developers evaluating pipeline architecture as the primary driver of AI system performance, the design is a concrete example of that thesis applied to a shipping product.
What Ships on the Device
Portable Computer packages the local model, inference engine, agent harness, tool sandbox, and connectors for Gmail, Outlook, Slack, and GitHub as a single installable system. Users choose between two local models at launch: Qwen 3.8 27B or PPLX 27B, Perplexity's post-trained variant tuned for its own harness. NVIDIA Nemotron 3.5 Lightning, an open 30B MoE model, is listed as coming soon. Bring-your-own model and inference server is also supported.
The OS-enforced sandbox restricts processes, filesystem paths, and network access for all code and tool calls. If the sandbox is unavailable, tool execution is disabled outright rather than silently downgraded — a deliberate fail-closed posture that matters for the finance, legal, and healthcare deployments Perplexity is explicitly targeting. The security implications of unsandboxed agent tool access have been well-documented, and this design reflects that awareness.
Context management required its own engineering. Qwen 3.8 27B advertises a 260K-token window but degrades in practice past roughly 100K tokens, so the harness keeps the system prompt and toolset small, loads specialized skills on demand, exposes connectors as compact CLI tools rather than full MCP definitions, and compacts stale context mid-run.
Escalation Gate and Benchmark Results
When a step requires live web access or frontier reasoning, the orchestrator halts and presents the user with exactly what context would leave the machine, after running a PII classifier over it. Only the approved context moves; the remote model returns text guidance and never receives direct access to local files, tools, or the full conversation thread. The system routes to one of 15+ cloud models on approval.
Perplexity published benchmark results across several evaluations:
| Benchmark | System | Score | Approx. Cost |
|---|---|---|---|
| Local Knowledge Work Bench (53 tasks) | Computer + PPLX 27B | 85.4% | Zero marginal (local) |
| Local Knowledge Work Bench (53 tasks) | Computer + Qwen 3.8 27B | 82.6% | Zero marginal (local) |
| Local Knowledge Work Bench (53 tasks) | Pi harness, same model | 77.6% | — |
| Local Knowledge Work Bench (53 tasks) | Hermes harness, same model | 74.0% | — |
| BrowseComp | Computer | 66.7% (51% less wall time, 70% fewer tokens vs. Pi) | — |
| BrowseComp | Pi harness | 50.2% | — |
| BrowseComp | Hermes harness | 43.9% | — |
| ParseBench-100 (visual document understanding) | Computer | 65.1% | — |
| ParseBench-100 | Pi harness | 34.6% | — |
| ParseBench-100 | Hermes harness | 13.9% | — |
| Terminal Bench 2.1 | Local only | 59.6% | ~$0 per rollout |
| Terminal Bench 2.1 | Local + adviser escalation | 73.0% | ~$0.415 per rollout |
| Terminal Bench 2.1 | Claude Opus 5 alone | 82.4% | ~$0.65 per rollout |
These are Perplexity's own published evaluations, not third-party replications. The harness comparison on the Local Knowledge Work Bench holds the model fixed and swaps only the agent scaffold, isolating the orchestration contribution. Terminal Bench 2.1 is the most useful figure for practical planning: adviser escalation narrows the gap to Claude Opus 5 by 13.4 percentage points at roughly 36% lower cost per rollout, but does not close it.
Hardware, Pricing, and Availability
The hardware requirement is the real barrier to entry. DGX Spark installs require the GB10 superchip, 128 GB of unified memory, and at least 1 TB of storage. Other Linux machines need DGX OS or Ubuntu on ARM or x64 with an RTX GPU carrying at least 24 GB of VRAM. The Qwen 3.8 27B orchestrator runs at 3-bit quantization, downloads as 17.4 GB, and requires 32 GB of RAM. The forthcoming Nemotron 3.5 Lightning uses 4-bit quantization, downloads at 19 GB, and requires 36 GB of RAM. Installation is via a standard apt repository add.
Portable Computer is available now for Linux on Pro, Max, Enterprise Pro, and Enterprise Max subscription tiers. Windows support follows in September; macOS is not on the roadmap. Only a single DGX Spark is supported per install at launch — clustering is a roadmap item, not a shipped capability.
The explicit cost-per-rollout figures published alongside benchmark scores indicate that Perplexity is positioning hybrid escalation as a tunable tradeoff rather than a binary cloud-or-local choice — a model that will put pressure on how agentic systems are priced and deployed. Whether the hardware gate proves an obstacle or a feature depends entirely on the customer's existing NVIDIA footprint.