Perplexity and Nvidia Launch Portable Computer Local AI Agent

Perplexity teamed up with Nvidia to introduce an offline artificial intelligence platform called Portable Computer. This newly released software allows professionals to run heavy processing tasks directly on their own hardware instead of relying on internet servers. By keeping the calculations on local machines, users avoid paying high token fees and guarantee their sensitive information never leaves their physical desktop.

The application is a highly optimized version of the original Perplexity Computer system. Nate Kupp, a representative for Perplexity, confirmed during a recent press conference that the offline software uses the exact same interface as the web version. While the local software handles most requests internally, it might ask permission to connect to a more capable cloud model for unusually difficult tasks. Users must grant clear consent before the software sends any text or data over the internet.

Shifting Computing Power to Local Machines

Nader Khalil, Nvidia's Director of Developer Technologies, explained that offline artificial intelligence is moving beyond a niche hobby. He noted that professionals now treat localized processing as a completely necessary tool for everyday work, mostly because it offers absolute data privacy. Companies want the benefits of intelligent software without uploading their corporate secrets to public servers. We observed similar privacy adjustments across the industry when Anthropic released Fable as a cheaper alternative featuring specialized data controls.

The internal architecture relies on an orchestrator and an agent harness to manage daily tasks. Users can choose between two main language models at launch. The options include Qwen 3.8 and PPLX, both scaling to 27 billion parameters. The engineering team plans to add support for Nemotron 3.5 Lightning soon. This addition brings a 30-billion-parameter Mixture-of-Experts model that activates 3 billion parameters per token.

Performance Results and Hardware Requirements

Early testing shows impressive numbers. On the Local Knowledge Work Bench, the PPLX 27B model achieved an 85.4 percent accuracy rating while consuming 678,000 tokens. The Qwen 3.8 setup scored 82.6 percent accuracy using 520,000 tokens. Both models outperformed competing systems like Pi and Hermes across tasks involving financial analysis, writing documents, and organizing files. The Portable Computer also finished first in the BrowseComp and ParseBench-100 tests while using the lowest number of tokens among all competitors.

Running this heavy software demands serious hardware. The application is currently available to Pro, Max, and Enterprise subscribers using Linux desktop setups. Users need an Nvidia RTX graphics card with at least 24GB of video memory. The software also runs on the Nvidia DGX Spark, a compact desktop built around the Grace Blackwell GB10 chip that features a central processing unit with 20 computing units and 128GB of unified memory. This heavy hardware reliance mirrors other industry movements, similar to how software capabilities expand when Google Pics introduces AI Pro Ultra image editor processing.

Windows compatibility is scheduled to arrive in September 2026. The Windows release will require the exact same 24GB video memory limit. By enforcing these high hardware requirements, the developers ensure the software runs fast enough to remain useful in professional environments.

Read More on TechRobust:

• CrowdStrike and OpenAI expand partnership to secure AI agents
• Lambda secures one billion private loan for Nvidia chips
• Nvidia leadership questions general machine intelligence milestones