Perplexity and Nvidia Launch Zero Token Cost Portable Computer AI Agent
Perplexity and Nvidia launch Portable Computer, a local AI agent for RTX systems with zero token charges, private workflows and optional cloud support.
Summary
Perplexity launched Portable Computer with Nvidia on August 25, 2026, packaging the Computer agent’s full stack into one local app. Tasks default to Nvidia DGX Spark or Linux PCs with an RTX GPU carrying at least 24GB VRAM, roughly GeForce RTX 3090 or newer, consume no billing credits, and require permission before cloud escalation. Pro, Max, Enterprise Pro and Enterprise Max subscribers get Linux access now; Windows follows in September, while Apple silicon has no roadmap. Qwen 3.8 27B and Perplexity-post-trained PPLX 27B are available, with Nvidia Nemotron 3.5 Lightning coming soon.
A DGX Spark demo ran a 27-billion-parameter Qwen model at full GPU utilization across 1099s and investment files, flagging unnecessary fees with credits at zero. Another analyzed a startup CSV locally, then posted results to Slack; Google Drive, Gmail and GitHub connectors are included. Compact command-line connectors, on-demand skills and self-verification reduce context pressure after Perplexity found Qwen 3.8 27B struggled beyond 100,000 tokens despite a 260,000-token claim. Always-on sandboxing disables the harness if unavailable. Cloud calls scan outgoing context for PII and disclose it; advisors return text only, without file or tool access.
On Perplexity’s planned open-source, 53-task Local Knowledge Work Bench, Computer with Qwen scored 82.6%, versus Pi’s 77.6% and Hermes’ 74.0%; PPLX reached 85.4%. BrowseComp scores were 66.7%, 50.2% and 43.9%, with Computer using 51% less time and 70% fewer tokens than Pi; multimodal scores were 65.1%, 13.9% and 34.6%. On Terminal Bench 2.1, local Qwen scored 59.6% near zero marginal cost, hybrid Claude Opus 5 scored 73.0% at $0.415 per task, and cloud-only scored 82.4% at $0.65. Company-run tests, weaker local reasoning and the hardware floor temper the privacy and cost case. Linked Sparks can run DeepSeek’s latest, GLM 5.2 or Nemotron Ultra.
Positives
- Local execution consumes no billing credits and keeps models, files and work on hardware the user controls.
- Computer scored 82.6% on the 53-task Local Knowledge Work Bench, while PPLX 27B reached 85.4%.
- Hybrid Claude Opus 5 escalation reached 73.0% on Terminal Bench 2.1 for an estimated $0.415 per task.
- PII scanning, explicit disclosure and text-only cloud guidance prevent remote advisors from accessing local files or tools.
- Slack, Google Drive, Gmail and GitHub connectors let local workflows interact with existing business systems.
Risks & concerns
- The 24GB VRAM minimum, roughly an RTX 3090, excludes most consumer PCs.
- Linux is the only supported operating system at launch, with Windows due in September and no Apple silicon roadmap.
- Local Qwen scored 59.6% on Terminal Bench 2.1, below the cloud-only model’s 82.4%.
- Perplexity produced the strongest benchmark results through its own evaluations, limiting independent validation.
- Qwen 3.8 27B struggled beyond 100,000 tokens despite advertising a 260,000-token context window.