Puppy One on NVIDIA.
Grace-Blackwell deskside supercomputing. PetaFLOP-class local inference and fine-tuning. Agent coming to these platforms; reserve your place.
Every NVIDIA Puppy, spec for spec.
For the geeky, token-hungry human: the full configuration matrix, with the local model sizes each machine can actually run on-device.
Puppy One Ultra S
DGX Spark · GB10
A petaFLOP on your desk. Personal AI workstation.
- Superchip
- GB10 Grace Blackwell
- CPU
- 20-core Arm (10× Cortex-X925 + 10× A725)
- GPU
- Blackwell · 5th-gen Tensor Cores · FP4
- AI performance
- 1 PFLOP (1000 TOPS) FP4 sparse
- Unified memory
- 128 GB LPDDR5x coherent
- Memory bandwidth
- 273 GB/s
- Networking
- ConnectX-7 · 200 GbE (cluster 2 units)
- Storage
- up to 4 TB NVMe
- Local models
- up to ~200B single · 405B across two units
Puppy One Ultra Max
DGX Station · GB300
Trillion-parameter local models. Datacenter on a desk.
- Superchip
- GB300 Grace Blackwell Ultra
- CPU
- 72-core NVIDIA Grace (Arm Neoverse)
- GPU
- Blackwell Ultra · FP4
- AI performance
- up to 20 PFLOPS FP4
- Coherent memory
- 748 GB (252 GB HBM3e + 496 GB LPDDR5X)
- GPU bandwidth
- HBM3e @ 7.1 TB/s
- Networking
- ConnectX-8 SuperNIC · up to 800 GbE
- Power
- Deskside · standard wall power
- Local models
- Trillion-parameter class · fine-tune 70B locally
Hardware specs are verified against NVIDIA's published figures for the named silicon; final configuration, memory, and pricing are confirmed at purchase. On-device model sizes and tokens/sec are estimates for int4/FP4-class local inference, not vendor figures.
The One Agent, pre-installed.
One Agent, pre-installed.Every NVIDIA Puppy ships with the One Agent already installed and configured - so it works the moment you open it.
Open the box
Your 🤫 Puppy powers on with the 🤫 One Agent already installed, configured, and updated. No setup-wizard marathon, no app store hunt - it is ready the moment you are.
Share your phone number
Your number is your identity key. It links your Puppy to the information bankers who already hold data against your profile - data that you, and only you, own. 24/7/365.
Sign in with what you already use
No new password to invent. Use the identity you already trust on this platform - the agent inherits your existing, secure login.
On NVIDIA, sign in with Google Account or Microsoft Account.
It's already yours, working
From the first minute the agent is on-device and on your side - reading in only what you consent to, ready to answer. Day 0, not day 30.
Meet your NVIDIA Puppy.
Reserve any configuration for any amount you choose, from $0.01, fully credited to your purchase and refundable anytime. We pre-install and configure the One Agent before it ships.
One is made by Hushh Technologies Corporation: private, sovereign AI that runs on what you own.