GPU-accelerated on the GPU you already have.
Running models locally usually means a setup project: a CUDA toolkit, a Python environment, a driver version that has to match. privstory is a native app with the acceleration built into the download. Install it, and it uses whatever your machine has.
One installer per platform, no toolkit
On macOS, Apple Silicon uses the Metal backend automatically. There is nothing to select and nothing to install alongside it.
On Windows and Linux the builds use Vulkan, which covers NVIDIA, AMD, Intel and integrated GPUs from the same installer. There is no separate CUDA build to pick and no driver toolkit to set up. If no usable GPU is found, the app falls back to the CPU rather than refusing to start.
How much memory you actually need
There is no single number, because privstory doesn't ship the models. A module brings its own chat model and image model, so the requirement is whatever that module's models need. A small module runs comfortably on modest hardware; a large one wants a lot of VRAM.
The useful detail is that the two models are never resident at the same time. Peak memory is set by the larger of the chat model and the image model, not by their sum, which is what makes a module with two substantial models workable on a single mid-range GPU.
What performance feels like
Text generation is the lighter half. On a recent Apple Silicon Mac or a discrete GPU, replies stream at roughly reading pace, which is the point at which the conversation feels live rather than batch-processed.
Image generation is the heavier half and the clearer test of your hardware. On a capable GPU a portrait or backdrop arrives in seconds. On integrated graphics it takes noticeably longer, and on the CPU fallback it is slow, usable to try the app out, but not what it's like on a GPU.
Nothing is offloaded to us
There is no hybrid mode where hard work quietly goes to a server. Every token and every pixel is computed on your machine, which is what makes the privacy and the absence of usage limits possible at all.
The consequence is worth stating plainly: your hardware is the performance ceiling. That's the trade you're making for a chat with no metering, no account, and no data leaving the device.
Common questions.
Do I need an NVIDIA GPU on Windows or Linux?
No. Every download is GPU-accelerated out of the box. macOS uses Metal; the Windows and Linux builds use Vulkan, so NVIDIA, AMD, Intel and integrated GPUs are all supported by the same installer, with no separate build or driver toolkit. The app falls back to the CPU if it cannot find a usable GPU.
How much VRAM do I need?
It depends on the module, since the module supplies both models. Because the chat and image models are never in memory at the same time, you need enough for the larger of the two rather than for both together.
Will it run without a GPU at all?
Yes, on a CPU fallback. Text generation is tolerable; image generation is slow. It is a way to try the app rather than the experience it is built for.
Does anything run on a server for performance reasons?
No. There is no offloading of any kind. The only network access is the one-time model download and, if you enter a license key, a lightweight license check.