Perplexity Teams Up with NVIDIA on Fully Local Agent Stack 'Portable Computer'

On Aug 25, Perplexity announced that 'Portable Computer' is available immediately to all Pro and Max subscribers, a local-first agent stack built in partnership with NVIDIA and optimized for DGX Spark. The entire runtime — orchestrator LLM, sub-agent LLMs, and agent framework — runs on local hardware without cloud dependency. CEO Arav Srinivas argues that in a compute- and power-constrained world, much agent inference must move to local devices.

Confirmed

• Available immediately to Pro and Max subscribers with DGX Spark integration, one-click local inference setup, and an optimized agent experience balancing privacy and cost
• Available models include Perplexity's post-trained PPLX 27B and Qwen 3.8B; OCR reaches SOTA performance in fully local use, which Srinivas says is critical for sensitive private documents
• Srinivas demonstrated a fully local agent runtime built on a 27B MoE, adapting the cloud Computer Harness to a local-first version orchestrated by the local model on DGX Spark; across 1,266 BrowseComp web-research tasks, Computer outperformed other local control frameworks
• To decide when a 27B local model should call a frontier model, Perplexity uses an 'advisor escalation' hybrid architecture: the local model runs the agent and escalates to cloud frontier models when stuck. Per ryanshrout, escalation decisions are made by a deterministic code orchestrator rather than an LLM; this hybrid agentic system delivered a 13.4% improvement on Terminal Bench
• On privacy, Perplexity says the coordinator asks user approval before routing to cloud frontier models, applies PII tagging to sensitive documents so they stay local, and only sends text instructions to the cloud
• Srinivas also shared the accompanying research blog 'Portable Computer on Spark'; per CNBC (relayed by dd32x), NVIDIA separately announced Perplexity will also run on its DGX Cloud infrastructure

Unconfirmed

• The 'upcoming Qwen 3.8 flash' mentioned by bakawolf123 has not been officially confirmed

Why it matters

This is an attempt to push the entire agent stack down to local hardware, achieving cloud-free agent orchestration with DGX Spark and small self-trained models, using a deterministic orchestrator to bound small-model judgment, while PII tagging and user-approval gates keep sensitive data local. Commenters like camerontstow argue that with progress in open models such as GLM 5.2 and DeepSeek, fully local agent stacks are rapidly becoming viable.

2026-08-25 ~ 2026-08-27 · 22 related posts

• NVIDIA and Perplexity Launch Portable Computer with Hybrid Agent Stack — ryanshrout · 2026-08-25
• Perplexity introduces hybrid orchestration for local privacy and cloud reasoning — perplexity_ai · 2026-08-25
• [source] Perplexity launches Portable Computer feature for Pro and Max subscribers — perplexity_ai · 2026-08-25
• Perplexity launches Portable Computer local agent stack for NVIDIA DGX Spark — NVIDIAAI · 2026-08-25
• [source] Arav Srinivas: Agentic inference must move to local hardware — AravSrinivas · 2026-08-25
• Perplexity launches local Portable Computer with PPLX 27B — testingcatalog · 2026-08-25
• [source] Perplexity demonstrates fully local agent runtime on DGX Spark — AravSrinivas · 2026-08-25
• Perplexity's Local Computer Harness Tops BrowseComp Benchmarks — AravSrinivas · 2026-08-25
• Perplexity Launches Portable Computer with Local Models on NVIDIA DGX Spark — testingcatalog · 2026-08-25
• Arav Srinivas: On-device models critical for sensitive docs with SOTA OCR — AravSrinivas · 2026-08-25
• Hybrid agent system closes local model gap on Terminal Bench 2.1 — AravSrinivas · 2026-08-25
• Perplexity hybrid agent system: local models with cloud escalation boost efficiency — AravSrinivas · 2026-08-25
• Perplexity releases research on Portable Computer on Spark — AravSrinivas · 2026-08-25
• Perplexity launches Portable Computer for local AI agents — camerontstow · 2026-08-25
• Perplexity Launches Portable Computer for Fully Local Execution — every · 2026-08-26
• Nvidia Partners with Perplexity AI to Run on DGX Cloud — dd32x · 2026-08-26
• Perplexity releases local-first Portable Computer, 27B model beats open-source harnesses — perplexity_ai · 2026-08-26
• Perplexity and Nvidia Partner for Local-First AI Platform — bakawolf123 · 2026-08-26
• Perplexity launches Portable Computer, a fully local AI agent runtime — pswider · 2026-08-26

3 near-duplicate retellings: inductionheads · AravSrinivas · ChrisUniverse