"On-device AI" gets used loosely enough that it's worth being precise about what it actually means for an AI copilot that watches your screen. Does the video feed of everything you're doing leave your desk, or not? For Violoop, the answer is local-first, not local-only — a distinction worth understanding before trusting any device with a live view of your screen.
-
Violoop's raw screen feed is processed on-device by a local model, not streamed to a server for routine work.
-
Violoop still sends heavier requests to the cloud through your own API key. Local-first means most traffic stays local, not that all of it does.
-
The tradeoff for precision over marketing language: claiming Violoop is "fully local" would be false. The honest number is roughly 90% fewer cloud tokens on typical use, not zero.
Reserve the Violoop AI copilot →
1. What "on-device" actually covers
Violoop runs a Qwen 8B model, quantized to Q4, directly on its own hardware: an RK3576 octa-core processor paired with a 26 TOPS NPU accelerator and 13GB of memory (8GB LPDDR4X plus 5GB of 3D-stacked memory). On that hardware, Violoop's own benchmark has the local model reaching roughly 53 tokens/second, about 2.7x faster than the same model running on a Mac mini M4 (roughly 20 tokens/second) in that same test. The local compute is real, not a marketing number attached to a chip that can't actually run the workload.
That local model handles the routine work for Violoop: reading what's on your screen, recognizing what app you're in, and routing most everyday requests without sending the screen feed anywhere. Raw screen data doesn't leave Violoop for this path.
2. Where the cloud still comes in
Local-first is not the same claim as local-only. The distinction matters because it changes what a privacy-conscious buyer should actually expect from Violoop. Heavier requests — the kind that need more reasoning than Violoop's on-device model can handle — route to a cloud AI model through your own API key; you bring the account. That routing cuts cloud tokens by roughly 90% compared to running everything through the cloud — routine, everyday use stays local on Violoop, while some fraction of harder requests still leaves the device.
"Local-first" here has one precise meaning for Violoop: the on-device model handles routine screen-reading and action-routing, and only requests that exceed its 8B-model capability route to the cloud.
3. Why the physical approval step is a separate claim from the privacy one
It's worth keeping two different Violoop design claims distinct, because they solve different problems. The local-first processing model is about where your data goes. A separate design — a dedicated STM32H563 security microcontroller that's the only component able to send keyboard/mouse commands to your computer — is about who can act on your behalf. Nothing executes without a physical key press. One is a privacy claim; the other is a control claim. Neither implies the other, and neither has been field-tested yet, since Violoop is still in Kickstarter-stage reservation, not shipping.
The choice comes down to this
If you want an AI copilot and treat "where does my screen data actually go" as a real question, not a marketing checkbox, here's the useful check — for Violoop or any other device: does it process routine requests locally by default, and is it explicit about what still goes to the cloud and why? Violoop's answer is yes on both counts. "Fully local" claims that don't account for harder requests are usually either wrong, or describing a much less capable model than the one doing your everyday work.
FAQ
Does Violoop send my screen to a server?
For routine, everyday use, the on-device model processes the screen feed locally and nothing is sent to a server. Heavier requests that need more reasoning than the local model can handle route to a cloud AI model through your own API key — a design targeting roughly 90% fewer cloud tokens than an all-cloud approach.
Is Violoop fully offline?
No — it's local-first, not local-only. The bulk of routine screen-reading and action-routing happens on-device, but it still depends on a cloud model (via your own API key) for harder requests.
Whose cloud account does the cloud portion use?
Yours. Violoop routes cloud requests through your own third-party AI API key rather than a Violoop-operated cloud service — so once a request leaves the device, it's subject to whatever retention policy your chosen provider has, not Violoop's.
Is local AI processing the same thing as the physical approval step?
No — they're separate design claims. Local-first processing is about where your screen data is analyzed. The physical approval step is a separate hardware gate (a dedicated security microcontroller) controlling whether an action can actually execute. A device could have one without the other — Violoop has both.
Sources
(Facts in this article are drawn from violoop.ai and violoop.ai/reserve/, verified 2026-07-16.)