Wherever it runs,it runs inside your boundary.
Score ships as a self-contained package rather than a service you call. Three ways to run it, each best for something different.
On-premises
Runs on customer-owned servers
- Installs on Intel Xeon, AMD EPYC or NVIDIA GPU servers you already own, bare metal or virtualized.
- Supports connected or air-gapped environments. Nothing resolves from the internet at install or run time.
- Ships as a self-contained Docker image with documented sizing guidance.
Best forControl and predictable capacity
Private cloud
Runs inside your cloud account and VPC
- Large-memory CPU and GPU instances from AWS, Google Cloud, Azure or Alibaba Cloud, in the region you need.
- No physical hardware ownership required.
- Traffic never leaves your account. Hybrid CPU and GPU layouts where a workload earns them.
Best forRegional availability and faster deployment
Data center
Runs in colocation or provider infrastructure
- Commodity CPU and GPU nodes a data center already knows how to rack, power and cool.
- Designed for capacity operators and single-tenant deployments, one endpoint per tenant.
- More throughput per rack and per kilowatt, on standard nodes with a long useful life.
Best forInfrastructure economics and partner delivery
Installation
Four steps to a serving endpoint.
The same sequence on every deployment option. Watch it run.
From a Docker image to a serving endpoint, in four steps
Install sequence
step 1 / 401Pull the image
02Install on your infra
03Start the service
04Point your agent at it
| Step | What happens |
|---|---|
| 1. Pull the image | One self-contained Docker image |
| 2. Install on your infra | Bare metal, VM or your cloud account |
| 3. Start the service | Load a model, expose the endpoint |
| 4. Point your agent at it | OpenAI-compatible, no code changes |
Benchmark your workload.
Send us a real workload. We run it on your models, your context lengths and your concurrency, and hand you the numbers either way.