A palm-sized inference terminal.
Plug it in. Load a model. Make it yours.
Concept render · pre-production.
Not a photo of a shipping unit.
Keep the whole loop on your desk.
Prompts and files never leave the device unless you send them.
Once a model is loaded, gpuvera runs with the network unplugged.
Buy it once. No per-token billing. The hardware is yours.
It sits next to your monitor, not in a server rack. Connect the terminal to your own network.
Open the local console in your browser and pull an open-weight model — chat, code, or vision. It stays on-device for next time.
Your laptop, phone, or another workstation talks to gpuvera over your LAN.
A palm-sized terminal for the open-weight models people run locally. Your devices share one box over your own network.
See the hardware targets ↗
These are pre-production units. Figures marked target are engineering goals, not independently verified measurements.
Every throughput number will ship with the model, quantization, and prompt length it was measured at.
Open-weight chat, code and vision models, sized to its on-device memory. Frontier cluster-scale models are not the goal. Exact sizes are targets on the specs page.
These are pre-production units. The images are concept renders, not photos of a shipping unit; final hardware may differ.
Join the waitlist for pricing, ship window, and refund terms as they firm up. No charge to reserve.
Join the waitlist for pricing, ship window, and refund terms.