Perplexity’s on-device agent, Portable Computer, is now part of the Windows app for PCs with an NVIDIA GeForce RTX or RTX PRO GPU carrying at least 24GB of VRAM. The model that actually runs there is PPLX 27B, Perplexity’s own post-trained variant of the open-weight Qwen 3.8 27B. Perplexity’s product page states plainly that “Qwen 3.8 27B is not available on Windows RTX PCs” — the base model that NVIDIA’s own launch post describes running there, without naming that distinction.
The third hardware surface in three weeks
Portable Computer first shipped Aug. 25 for NVIDIA’s DGX Spark on Linux, then extended to Linux RTX PCs; Sept. 14 adds Windows RTX and RTX PRO Workstation GPUs as a third surface. On Windows, it works inside the existing Personal Computer app, launched July 28, and it does not replace either that cloud app or the separate cloud-based Perplexity Computer that launched Feb. 25: turning on local inference runs PPLX 27B entirely on the user’s own hardware instead. It requires a paid Pro ($20/month) or Max ($200/month) subscription, on individual or enterprise plans. Before any step goes to a cloud model, a PII classifier screens the context, shows the user what would leave the device, and waits for explicit approval. The local orchestrator reaches Gmail, Outlook, Slack, GitHub and locally installed desktop apps through local MCP servers running on the Windows machine.
One model, three scores, and two different search engines
Perplexity’s research blog backs the release with results from its own 53-task “Local Knowledge Work Bench” — run on DGX Spark, not on a Windows RTX PC. There, Computer scored 82.6% with the base Qwen 3.8 27B and 85.4% with PPLX 27B, against 77.6% for Pi and 74.0% for Hermes running that same base model in their own harnesses. Perplexity’s own text adds a detail that changes what those numbers show: Computer used Perplexity’s search infrastructure in that comparison, while Pi and Hermes used Brave, their recommended provider, and deep-research tasks that depend on search quality make up 37.7% of the 53 tasks. The same research post says the on-device model “begin[s] to struggle beyond 100K tokens,” under half of the 260K-token context window Qwen 3.8 27B advertises.
An unconfirmed report puts NVIDIA on both sides of the launch
SiliconANGLE reported at the Aug. 25 launch that NVIDIA was “weighing an investment in the startup that could value it at over $30 billion,” and had “floated the idea of licensing Perplexity’s technology and hiring key employees.” Neither company has confirmed this, no second outlet has corroborated it, and it rests on SiliconANGLE’s own unnamed sourcing. Treated as a report rather than a fact, it still reframes NVIDIA’s launch post as coming from a company that may be about to invest in Perplexity, on top of selling it GPUs.