Perplexity AI launches Portable Computer on-device AI agent
SiliconANGLE Maria Deutscher ● Covered by 4 sources
Perplexity launched Portable Computer, an AI agent that runs on Nvidia desktop hardware. It brings its cloud tricks on-device, with guardrails and a fallback to the cloud.
Based on reporting by SiliconANGLE, Maria Deutscher — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Perplexity AI has pushed its agent software closer to the machine. The company today introduced Portable Computer, an on-device version of its cloud tool that runs on desktops built with Nvidia silicon.
That matters because Perplexity is not treating this like a toy demo. The agent is meant to handle multi-step work, including building web apps. When a job gets too heavy, it breaks the task into smaller pieces and hands them to separate subagents, the same basic playbook Perplexity uses in the cloud version it launched in February.
The first target is Nvidia’s DGX Spark, a desktop system with a Blackwell-based graphics card, a 20-core CPU and 128 gigabytes of memory. On launch, Portable Computer uses the open-source Qwen 3.8 27B model, and it also supports Perplexity’s own PPLX 27B variant tuned for the same machine. Perplexity says it added a multi-token prediction system to speed up prompt handling.
There’s a catch, though. Qwen 3.8 27B has a 256,000-token context window, but Perplexity found it starts to struggle above 100,000 tokens. So the company built in context compaction, which summarizes long prompts to pull them back under that limit.
The software also comes with packaged skills for online research, data science and coding, plus a check that verifies response accuracy. If the local model can’t finish a task, it can ask a cloud-based neural network for help. It can also work with services such as GitHub, but only after asking permission, and it does not expose on-device files to those tools. Perplexity says it also uses a sandbox and blocks unauthorized network connections.
An update will expand support beyond DGX Spark to Windows machines with Nvidia RTX graphics cards. Perplexity also plans to add Nvidia’s Nemotron 3.5 Lightning model, a 30-billion-parameter mixture-of-experts system built for lower-powered devices.
My take — AI-written commentary, not fact-checked reporting
This is the obvious direction: keep the agent close to the data, then phone home only when needed. Perplexity is selling convenience, but the real pitch is control — and Nvidia gets to be the toll booth on both sides. Open models are winning the unsexy parts of AI, which is exactly where the useful stuff lives.
Read more about this at: SiliconANGLE