Caleta Private AI Appliance
by Caleta IT Solutions Limited
Frontier open-weight LLMs served privately in your own Azure subscription.
Private AI does not need to cost thousands a month. The Caleta Private AI Appliance runs open-weight language models on Azure spot capacity, and tells you where to run them: the cheapest regions for your VM size, ranked by measured eviction rates. Spot economics without the roulette.
One Azure VM becomes a private model server inside your own tenant. Deploy and SSH in, or open the VM's serial console in the Azure portal. Either way, the menu opens on login: pick a model from the catalogue and serve it through a bundled chat UI or an OpenAI-compatible API. Nothing leaves your subscription.
Key Solution Features:
- Spot placement built in: a live ranking of the cheapest regions for your VM size, scored against real eviction rates over 24-hour, 7-day and 30-day windows. Azure does not publish this data; we measure it. Pick a region on evidence instead of guessing.
- Serving in about five minutes: deploy to first answer, model download included. We timed it. There is no stack to assemble.
- Downloads at hyperscale speed: the download path is tuned for the fastest possible pull from Hugging Face: multi-gigabit, measured at 30 Gbit/s on the largest sizes. Models that take an hour to fetch elsewhere land in minutes.
- Park it overnight: deallocate when idle and the only standing charge is the OS disk. On restart, or after a spot reclaim, the box re-downloads its model and serves again unattended.
- Honest numbers: where we publish a speed, we measured it ourselves on this exact image and dated it. Where we have not measured a pairing, the catalogue says so rather than guessing. Serving uses the engine's own defaults, not tuning we invented. The current models and numbers are at https://caleta.io/appliance/models/
- The menu stays current: new models publish to our catalogue service. Pull them onto a box you already deployed, no rebuild, no new image. It updates only when you ask it to.
- Two architectures: x86 sizes detect the GPU and serve through CUDA on NVIDIA or ROCm on AMD automatically. Azure Cobalt Arm sizes serve on CPU with one large memory pool. Frontier mixture-of-experts models run with no GPU at all, at single-user speeds. Yes, really.
- Private by default: chat UI and API bind to localhost, the host firewall is on, and a private API key is minted on your VM, never baked into the image, and kept across reboots. Publish on :443 with an automatic certificate only when you choose to, with optional source IP restriction.
- Expert-friendly: every serve setting is visible and editable at serve time. The menu is the easy path, not a cage.
- No bundled weights: models download under their own licences, direct from Hugging Face. The image ships engines and tooling only.
Who it is for:
Enthusiasts, developers and teams who want open-weight models with full data sovereignty, without building the serving stack themselves.
Prerequisites:
You need an Azure subscription with vCPU or GPU quota for your chosen VM size. Setup help is in the online guide.
Getting started:
Deploy, then SSH in, or use the Azure portal's serial console if you would rather not install a client. Either way the menu opens on login. Pick a model and serve it. The trial model downloads in about a minute and the first chat is minutes away.