pipelines
A pipeline goes live whole, or not at all.
If any stage fails, every machine reverts.
Fleet Management
Cascadia finds the machines with room and places the model for you: a copy on each, or one model split across several. Add machines with a link, see usage, and any change that fails rolls back on its own.

One 35B model split across two machines, and a whole model on each of the other six.
API endpoint
3 models on 8 machines
01
Fleet Management checks it against every machine’s free memory.

Qwen2.5 7B
7B · int4
4.5 GB
Phi-4 14B
14B · int4
8.1 GB
Qwen2.5 32B
32B · int4
17.8 GB
Qwen3.6 35B
35B MoE · int4
21 GB
largest machine · 18.4 GB free
02
It picks the machines with the most room. Prefer to choose them yourself? Switch to manual.

Qwen3.6 35B
35B paramsMoE · int440 layers21 GBL1
L20
L21
L40
stage 1 · L1–20
lab-05
10.5 of 18.4 GB
stage 2 · L21–40
lab-06
10.5 of 17.9 GB
03
Stages commit together, then one endpoint serves it.

Deploy Qwen3.6 35B
Download
Load
Staged
Commit
Verify
holding 1:42
Serving
lab-05
stage 1
lab-06
stage 2
API endpoint
healthy
http://198.51.100.15:9110/v1
qwen3.6-35b
Built with Intel for the AI PCs you already own. Cascadia runs every model on OpenVINO, Intel’s own AI runtime. Models are compressed to fit an office PC’s memory and run on the processor, graphics and NPU built into Intel Core Ultra.
~30 tok/s
An 8B model on one Intel AI PC, with speculative decoding.
14B on one PC
4-bit weights fit a 14B model in a 32 GB office machine.
CPU, GPU and NPU
OpenVINO uses every engine in the chip, including the NPU most software leaves idle.
the hand-off
IT doesn’t have to touch the computer. Send the person a link; they click one button; you watch each step land.
Nothing to unblock in the inbox
No files or commands in the email, so it clears mail filters.
One link, one machine
Each link adds a single device and expires in seven days.
A whole room at once
Doing it yourself? One batch installer enrols up to 500 machines.
usage and status
Tokens, requests and success rate over any range, by model, device or API token, with a year of daily activity. Status checks the fleet every 15 seconds and lists what needs attention.
automatic rollback
pipelines
If any stage fails, every machine reverts.
model switches
The previous model stays on disk, ready to take over.
updates
Fleet Management keeps a new version only if it passes a self-test.
cascadia desktop
Desktop is the private AI app for every employee. It runs on the models you deploy here, and nothing anyone types leaves your network.

FAQ
Most teams don’t. It’s built for the IT team you already have: Cascadia places models and rolls back failed changes on its own, and we set up your pilot with you.
Yes. Placement is automatic by default. Switch to manual to pick exact machines, set how a big model splits, or choose CPU, GPU or NPU.
No. It reads health and usage: which machines are up, what each serves, request and token counts, latency. Prompts and replies stay on the machines that serve them.
One at a time. Each machine can keep several models on disk and switch between them without downloading again. To serve more models at once, give each its own machines, or split one across several.
Yes. For an air-gapped fleet, Fleet Management and a certificate authority run on your own network in place of the hosted service, so the machines never need to reach the internet.
An Intel Core Ultra PC running Windows 10 or 11, or Ubuntu 22.04 or later, and administrator rights to run the installer once. Free memory decides what a machine can hold on its own: a 32 GB machine fits models up to about 14B, and Fleet Management splits bigger models across several machines.
Open-weight models such as Llama, Qwen and Phi. Fleet Management checks each one against every machine’s free memory, then runs it whole on one machine or splits it across several, up to 1T-class mixture-of-experts models.
Per device, never per token, so the bill stays flat as usage grows. Talk to us for pricing for your fleet.
We set up a pilot on machines you already own, and your team deploys from Fleet Management from the first day.
See it run on hardware like yours.
Book a demoCheck the engineering first.
See our research