Frontier AI without accelerators.
Serve frontier models at their original precision on off-the-shelf hardware, or on the servers you already own, without GPU cards or AI accelerators. Piccolo makes that possible by building the runtime around your specific model and hardware.
Piccolo Rack
Frontier AI for your business
Rewritten for your model
Generic runtimes are built to run any model on any machine. Once the model is known, everything can be rewritten for it, from the order of every computation down to the network between machines.
That is what lets modest hardware serve a frontier model at its original precision.
- The order of every computation
- The integrated-GPU kernels
- The operating system and drivers
- The network between machines
Two ways to get it
-
A ready Rack
A ready machine built from off-the-shelf hardware. Plug in power and a network connection, and you have an OpenAI-compatible endpoint inside your network. We keep the runtime current as models change.
-
On your own servers
Already have servers without accelerators? Tell us the model you want to serve and the machines you have, and we bring the same engineering to your stack.
What it lets you do
-
Give every customer their own frontier model.
Dedicated to them and kept in the region you choose, for every customer, not just the biggest.
-
Keep your data in-house.
Your code, contracts and customer data stay inside your network, served by the same frontier models the cloud offers.
-
Serve agents and long documents.
Long context stays loaded for coding agents and assistants that keep returning to the same material.
The first Rack
Forty off-the-shelf mini PCs working as one machine to serve a frontier model. No GPU cards, no AI accelerators. The story of our first build.
Follow the build on LinkedInTell us what you want to serve.
The model, the hardware you have, and what's getting in the way. We'll bring this engineering to your stack.