Skip to content
Code you can read · offline · your files stay yours

Local AI tools you own.

Velquaria builds AI software that runs on your own computer. Your files stay on your disk, nothing is rented, and it keeps working when the internet doesn’t.

Same model, same GPU, same answerRTX 4000 Ada · 20 GB
How does a 125B MoE model fit on a 20 GB GPU?
llama.cpp16–18 tok/s

A 125B mixture-of-experts model only uses a few experts per token. Strata keeps the experts in system RAM and the rest on the GPU, so a 20 GB card holds what each token needs, and the model answers at reading speed — on your own machine, with nothing sent to the cloud.

—
Velquaria Forge + Strata60–80 tok/s

A 125B mixture-of-experts model only uses a few experts per token. Strata keeps the experts in system RAM and the rest on the GPU, so a 20 GB card holds what each token needs, and the model answers at reading speed — on your own machine, with nothing sent to the cloud.

—
Replayed at the measured generation speeds. Qwen3.8-Flash-Next 125B MoE, RTX 4000 Ada 20 GB + 125 GB RAM.
Why local

Your machine, your data, your tools.

Your data stays home

Code, documents and chats live on your disk. Nothing is uploaded to train someone else’s model.

No subscriptions

No monthly bill for tokens or seats. The models run on hardware you already paid for.

Works offline

On a plane, behind a firewall, during an outage. Your tools don’t depend on someone else’s servers.

Code you can read

Code you can read, audit and keep: Velquaria Forge and Velquaria Studio are source-available (FSL-1.1), and each version becomes open source (Apache-2.0) after two years. If we ever stop, your copy keeps working.

Cloud models stay optional, with your own keys — and a local model never quietly falls back to the cloud.

Early access · free

Get the tools before everyone else. Plus a free field guide.

  • The guide, right away: which models fit which GPU, the settings that run a 125B model at 60–80 tok/s on 20 GB, the rules that make local coding agents reliable (with our measurements), and a sandbox checklist.
  • Early access to Velquaria Forge, then Studio, before the public release.
  • One short email when something real ships. Nothing else.

Free. The list runs on Substack (velquaria.substack.com); unsubscribe in one click. Sender: Velquaria, Italy — company details coming soon. Privacy