On-Premise
Deploy Xilos in your own data center or private cloud with full data sovereignty. On-Premise is air-gapped capable, has no external dependencies, and gives you complete control over the platform. This is the final phase of the three-phase journey.
What This Means for You
With On-Premise, the Xilos platform runs entirely on infrastructure you control. No data leaves your network. No external API calls are required if you self-host models. You own the hardware, the network, and the licensing. Xilos provides the software, deployment artifacts, and support.
Hardware Requirements
| Component | Minimum | Recommended |
|---|---|---|
| Application servers | 2 × 4 vCPU, 8 GB RAM | 4 × 8 vCPU, 16 GB RAM |
| Database server | 4 vCPU, 16 GB RAM, 200 GB SSD | 8 vCPU, 32 GB RAM, 500 GB SSD |
| GPU for internal SLM | 1 × NVIDIA GPU (16 GB VRAM) | 2 × NVIDIA GPUs (24 GB VRAM each) |
| GPU for fine-tuned models | 1 × NVIDIA GPU (24 GB VRAM) | 2 × NVIDIA GPUs (40 GB VRAM each) |
| Load balancer | Hardware or software | Hardware with HA |
Info: GPU requirements vary based on the models you intend to host. Xilos provides sizing guidance based on your expected query volume and model selection.
Installation
Obtain a License
Contact Xilos to obtain an On-Premise license and deployment package. The license is tied to your organization and includes a defined number of seats and models.
Prepare Your Infrastructure
Provision the required hardware or virtual machines. Install a supported operating system and container runtime. Set up a PostgreSQL database.
Deploy the Platform
Use the provided deployment scripts to install the Xilos application, database schema, and supporting services. Xilos provides installation documentation and remote support.
Configure Networking
Set up internal networking, load balancers, and DNS. If air-gapped, ensure all model weights and dependencies are available locally.
Import Models
Import your fine-tuned models and base model weights into the local model registry. If you previously fine-tuned models in Managed SaaS or Private Cloud, export and import them here.
Validate and Go Live
Run smoke tests, verify routing and restrictions, and confirm audit logging. Xilos support assists with validation.
Licensing
- Perpetual license — The license is issued for a defined term (typically annual) and includes software updates during the term.
- Seat-based — Licenses are scoped to a number of users and API keys.
- Model hosting — The license includes hosting of the internal SLM and a defined number of fine-tuned models.
- Air-gapped activation — Offline activation is supported for air-gapped environments.
Warning: On-Premise licenses must be renewed annually. Expired licenses result in read-only mode — existing configurations remain accessible but new queries are not processed. Contact Xilos for renewal.
Air-Gapped Operation
On-Premise supports fully air-gapped operation:
- No outbound internet connections are required.
- All model weights, dependencies, and platform components are installed locally.
- Software updates are delivered as offline packages.
- The internal SLM and fine-tuned models run on local GPU hardware.
Info: For air-gapped deployments, you must self-host all LLM models. OpenAI, Anthropic, and other cloud-based providers are not available. Use open-weight models like Llama or Mistral, or your own fine-tuned models.
When to Choose On-Premise
- Full data sovereignty — All data must remain within your physical premises.
- Air-gapped networks — No internet connectivity is available or permitted.
- No external API dependencies — All models must be self-hosted with no cloud provider calls.
- Maximum control — You need full control over hardware, network, and software stack.
Security
- Encryption — Full control over encryption keys and TLS certificates.
- Network isolation — The platform runs entirely within your network perimeter.
- Audit logging — All events are logged locally. Forward to your SIEM via CEF, Syslog, or JSON.
- Authentication — Integrated with your identity provider or use Xilos native authentication.
See Security Model and Governance for more details.