Quick Facts
- Portable Computer runs entirely on Nvidia DGX Spark desktops or RTX GPUs with at least 24GB of VRAM, with no token charges for local work.
- The system scored 85.4% on Perplexity’s 53-task Local Knowledge Work Bench using its post-trained PPLX 27B model, beating open-source alternatives.
- Available now on Linux for Pro ($20/month) and Max ($200/month) subscribers; Windows support arrives in September, with no macOS plans announced.
Perplexity and Nvidia on Aug. 25 launched Portable Computer, an agentic AI platform that runs entirely on local hardware. The product eliminates per-token billing for any work completed on-device and is the most direct challenge yet to cloud-dependent AI agent pricing.
The platform packages Perplexity’s full agent stack — model, inference engine, planner, tool router, task queue, and local search index — onto hardware users already own. It reads local files, parses PDFs, runs shell commands, browses the web, and executes code inside an OS-level sandbox. Documents never leave the machine unless the user directs otherwise.
How It Works
The orchestrator and planner run on-device and handle as much of each task locally as possible. When a task requires frontier reasoning, the system asks permission before routing work to one of 15-plus cloud models. At launch, users can run Qwen3.8-27B or PPLX 27B, Perplexity’s post-trained variant. Nvidia’s Nemotron 3.5 Lightning is coming later, and users can bring their own models and inference servers.
Perplexity built a deliberately minimal agent harness after finding that models advertising 260,000-token context windows begin to degrade beyond 100,000 tokens. The company converted popular connectors, including Gmail and GitHub, from token-heavy MCP servers into compact command-line tools. If the OS-level sandbox is unavailable, the harness shuts itself off rather than running tools with full user permissions.
For inference, Perplexity uses vLLM and jointly optimized the model and orchestration, rather than dropping a language model into a generic agent framework.
Hardware and Pricing
The primary target platform is Nvidia’s DGX Spark, a desktop system built on the GB10 Grace Blackwell superchip. It carries a 20-core Arm CPU, 6,144 CUDA cores, 128GB of unified memory at 273GB/s, and roughly one petaFLOP of FP4 compute. The Founders Edition price is $4,699 after a February 2026 increase from $3,999.
Any Nvidia RTX GPU with at least 24GB of VRAM also qualifies. Perplexity’s VP of engineering described that threshold as the floor needed to deliver a solid experience while keeping the platform broadly accessible. An RTX 3090 or newer meets the requirement.
Portable Computer is included in existing Perplexity subscription tiers. Pro costs $20 per month, or roughly $17 per month billed annually. The Max tier runs $200 per month, or about $167 per month on an annual plan. Enterprise Pro and Enterprise Max subscribers also get access.
Performance Numbers
On Perplexity’s 53-task Local Knowledge Work Bench, which spans deep research, financial analysis, and document creation, Computer running Qwen3.8-27B on a DGX Spark scored 82.6%. The open-source Pi harness on the same model scored 77.6%, and Hermes scored 74.0%. Switching to PPLX 27B pushed Perplexity’s score to 85.4%.
Nvidia noted that the hardware scales beyond a single unit. Two DGX Sparks connected over shared memory can run frontier open models including DeepSeek’s latest release. Four units can run GLM 5.2 or Nemotron Ultra.
Business Implications
Perplexity originally launched its Computer platform on Feb. 25, 2026, as a cloud-hosted agent workspace running across 19 AI models. The company, valued at $20 billion, moved Computer into the enterprise shortly after, putting it in direct competition with Microsoft, Salesforce, and established software vendors.
Portable Computer extends that push by removing the recurring token cost that makes high-volume agentic workloads expensive at scale. For software founders running internal automation, document processing, or research pipelines, shifting that compute to owned hardware changes the unit economics of AI-assisted work. The Linux release is live now. Windows support is scheduled for September 2026.