Your AI.
On Your Premises.
Under Your Privilege.

Privilege, sovereignty, and predictability. With on-premise, no prompt or completion ever leaves your network; you control every change window and every model upgrade. You also stop paying the silent cost of vendor changes — for example, Anthropic’s Opus 4.7 tokenizer change in April 2026 produces up to ~35% more tokens for the same input, an effective price increase the customer didn’t agree to. None of that happens on a unit you own.

A single Spark comfortably runs Qwen3.6-32B or Gemma 4 31B at FP8 for interactive use by a small team, plus 70B-class models in batch / RAG mode. Honest caveat: decode speed on the largest open models is bandwidth-bound (~273 GB/s LPDDR5x), so single-user 70B chat is in the low single-digit tokens/sec range. The sweet spot for an AU SME is 7B–35B class for chat, with larger models reserved for overnight batch.

It depends on concurrent users and workload. We size on assessment. For HA we recommend two Sparks linked over 200 GbE — combined 256 GB unified memory unlocks ~405B-class models at FP4 and gives you live failover. Beyond two units, we usually point clients at a DGX Station / H200 server instead.

We are deliberately model-agnostic. The Model Lab on this page lets you filter by industry, licence and Spark size. For regulated SMEs we typically recommend Qwen3.6, Gemma 4 or Phi-4 Reasoning (all Apache 2.0 / MIT). We flag Llama 4 (Llama Community Licence — has EU geofencing and use-case restrictions) and DeepSeek (foundation lab geopolitics) explicitly so the buyer can self-select.

On-premise inference keeps you on the right side of every one of those frameworks because the regulated data never traverses a third-party service. We document the controls (encryption at rest with LUKS, network egress filtering, Open WebUI RBAC + audit log to your SIEM, OEM-supported hardware) and provide an APP 1.7-1.9 transparency artefact for the 10 Dec 2026 deadline. See our /solutions/healthcare and /solutions/professional-services pages for the industry-specific framing.

We run a quarterly health review, manage a controlled model-update mirror, and coordinate OEM ProSupport tickets for hardware. Crucially, the box does not push outbound to the internet — we pull updates through a controlled mirror on your schedule, so you never face a surprise model deprecation or tokenizer change.

Yes. Open WebUI speaks OIDC for SSO into Entra ID, Okta or Keycloak, and supports MCP (Model Context Protocol) tools out of the box. We commonly wire SharePoint, OneDrive, SQL, and ticketing systems via MCP servers so the model can read what your team already has access to — under the same RBAC.