The 24GB Line Is the Actual Content of This Announcement
"Local agent" sounds like a philosophy until the hardware requirement turns it into a purchasing question: an RTX 3090 or better, with no less than 24GB of VRAM. Per Steam hardware data, roughly twice as many users own a premium NVIDIA card that falls short of that bar as own one that clears it, and business laptops and ordinary office desktops sit much further below. So the group this serves today is well defined but not large — developers, researchers and designers who already happen to own a big-VRAM card. For enterprise procurement, that line means "give everyone a local agent" is not a budget question right now; it is a hardware refresh question. This site covered Perplexity's Hybrid Compute on Mac on September 2 and its open-sourced local inference engine Lily on September 4 — an engine that serves exactly one model on Apple silicon only. Extending the line to Windows and RTX keeps the same shape: **a deliberately narrow hardware-and-model combination traded for predictable local performance.** The price is always the same one: you need the hardware first.
One Product Decision Worth Copying: Escalating to the Cloud Requires Asking
For anything marketed as local-first, the real risk was never the local part — it is when the thing quietly reaches for the cloud. What Portable Computer does here is worth writing down: when current information or heavier reasoning is needed it can escalate to a frontier cloud model, but **it asks first and screens the outbound payload for personal data**. Code and tools run in isolated environments, with controlled access to files and connected apps. Making that step an explicit confirmation rather than a silent fallback is the right default for this category. It also marks the boundary on the credits claim: "does not consume Perplexity Computer credits" holds only while the task actually completes locally. Once it escalates to a cloud model, that no longer applies.
Confirm Which Local Model You Are Getting Before You Install
Two official sources contradict each other here, which is worth noticing before you start: Perplexity's own product page states that Windows machines get PPLX 27B as the local model and explicitly says Qwen 3.8 27B is "not available on Windows RTX PCs," while NVIDIA's launch blog names Qwen 3.8 27B as the model the app sets up. There is no third-party clarification of the discrepancy yet. Go by whichever model the app actually downloads — which also determines whether you can reproduce any performance figures other people publish. Everything else is clean enough: it installs through the Perplexity app in the Microsoft Store, comes with Pro, Max, Enterprise Pro and Enterprise Max subscriptions rather than being sold separately, and DGX Station support was announced at the same time. If you are already on one of those plans and happen to own a 24GB card, this update costs nothing to try. Everyone else can note the threshold and wait for the next round of hardware or models to bring it down.
via: NVIDIA blog, VentureBeat, Cryptonomist