AI-first hardware, loaded with Agent One, built to handle the tedious work while you stay close, private, and in control.
What your Puppy is
Frontier power, owned by you.
Agent One is already on board.
The machine works hard. You stay in control.
It started with a puppy
Rio made the spec simple: loyal to one family, quick to help, careful with the house, never selling you out.
Every Puppy One is built to Rio's standard. That's not a metaphor, it's the spec.
In August 2021, 🤫 was founded and Rio Grande came home. One became the company; the other became its conscience.
Ready, across your Apple devices
Puppy gives Agent One a home base, then keeps the devices you already use close to the same private brain.
Apple, Mac, macOS, iPhone, iOS, iPad, iPadOS, Apple Watch, watchOS, Apple TV, tvOS, Apple Vision Pro, visionOS, Siri, and Apple ID are trademarks of Apple Inc. Hushh Technologies Corporation (🤫) is an independent company and reseller; this page is not produced, sponsored, or endorsed by Apple Inc. 🤫 Private Agent One works alongside Apple software and is not integrated into Apple's operating systems.
Choose your One
Pick the machine that fits your life. Change it anytime before you reserve.
Complete the pack
Start with Agent One, add Puppy for power, add Tag One when you want the whole family close.
Your private agent starts on the device in your hand. Upgrade when you want more power.
A personal supercomputer with Agent One already on board, ready for the tedious work.
Your supercomputer, your agent, and Tag One together, close to the people and things you love.
Plug in and your Puppy joins the network from anywhere, a garage, a warehouse, the edge of the grid. Best-in-class networking so your supercomputer is always reachable and always yours.
Platform power
Choose a platform, see the Puppy it unlocks, then open only the details you need.
Apple
Mac Studio · M5 Max
The bandwidth-per-dollar champion. 70B-class models resident on your desk.
Full specifications
Agent live
The $899 door to owning your agent. Apple's first 2nm chip, always on, on your shelf.
Agent live
The home mainstay. 30B-class models resident, and it clusters when you outgrow it.
Agent live
The bandwidth-per-dollar champion. 70B-class models resident on your desk.
Agent live
Frontier-class open-weight models, entirely in your house. The garage-grid node.
Agent live
The phone becomes your agent. Pocket-sized, always with you.
Agent live
128GB-class on-device AI. The personal supercomputer you own.
Agent live
A whole RIA practice, configured, with onsite onboarding.
Agent live
Software only, self-serve, cancel anytime. The easiest way in.
Specs are based on partner-published figures and Puppy estimates where labeled. Final configuration, memory, pricing, and availability are confirmed at purchase.
The whole idea
The best hardware on earth, loaded and ready, owned by you. No setup, no lock-in, no one selling you out.
A penny, or as much as $69,696.96 - fully credited, refundable anytime, and enough to prove a real human is behind the order. Your card details go to our payment processor only; 🤫 never sees or stores them.
Explore Hussh
Puppy is simple to reserve and deep to grow into. Choose where to go next.
Power explorer
Memory, not compute, is the scarce commodity. Every rung of this ladder is a unified-memory-per-dollar decision, and the honest comparison is the same model at three watt classes — on battery, on a shelf, and on a circuit.
Studio in focus
Mac Studio, M3 Ultra
Explore full specifications
Workstation-scale private intelligence that travels
What this does not do. Not the raw multicore leader, and no CUDA. Unified memory is shared by the CPU, GPU, OS and applications — it is not 128GB of dedicated video memory.
128GB of local AI in about three and a half pounds
What this does not do. Runtime compatibility must be workload-tested. The x86 unified-memory path is younger than CUDA and than Metal, and support varies by framework.
Full-power Blackwell CUDA without workstation pricing
What this does not do. A mobile 5090 is not a desktop 5090, and sustained power matters more than the model number — a higher-TGP 5080 can beat a power-limited 5090. Compare sustained performance, never burst.
The compact native-CUDA agent node
What this does not do. Not a general-purpose desktop. Priced against memory scarcity, and street pricing has moved with the DRAM market rather than with the spec.
Large local model capacity without a rack
What this does not do. The largest memory configurations are supply-constrained and the ceiling has moved. Configuration is confirmed at order, not assumed from the tier.
96 cores and a professional Blackwell foundation that scales
What this does not do. The exact maximum bill of materials is quote-only. Multi-GPU totals are aggregate VRAM across PCIe devices, not one pooled accelerator.
Datacenter-class coherent AI on a desk
What this does not do. Headline FP4 throughput figures assume sparsity. Capacity to run a very large model is not a statement about how fast it will answer.
Memory capacities, bandwidths and power figures on this page are the manufacturers' published specifications, not measurements we have taken. We publish no throughput, tokens-per-second or benchmark score for any rung, because we have not yet completed a Compute Passport on it - and a performance number without a test behind it is a guess with a decimal point. Configuration and pricing are confirmed at order.