Distributed AI Compute,
Powered by Everyone.
Horizon turns compatible machines into a coordinated compute network for AI inference. One control plane for your hardware, models, and workloads.
Operate compute,
not a collection of machines.
See workers, deployments, model runtimes, and health in one focused workspace. The preview below uses demonstration configuration based on Horizon's current MVP concepts.
Compute overview
Connected resources and active workloads.
A control layer for fragmented compute.
AI developers often have hardware in several places: personal workstations, lab machines, development PCs, and unused CPU or GPU capacity. Horizon gives that hardware one operational boundary for inference.
Different operating systems, architectures, and resource profiles.
Register hardware, target deployments, and observe runtime state.
The current MVP is intended for machines you own or explicitly control.
AI workloads need somewhere to run.
Cloud GPU infrastructure can be expensive. Personal hardware is often underused. Heterogeneous machines are difficult to operate consistently.
Compute is fragmented
Resources live across laptops, workstations, and lab machines with different capabilities.
Operations are scattered
Model runtime state, hardware health, and deployment configuration should not live in separate scripts.
Cloud is not always the answer
For trusted teams and local workflows, existing hardware can be the practical starting point.
A small, explicit path from machine to inference.
Connect compute
Install and run the Horizon Worker Agent on a compatible machine.
Register hardware
The worker reports hardware and availability to the Control Plane.
Deploy a model
Choose a model and target compatible compute for the workload.
Run inference
The model runtime serves requests on the selected worker.
Observe state
Track worker status, deployment state, and runtime health.
Different machines.
One operational view.
Horizon is designed around heterogeneous, trusted compute. Hardware stays where it is; the Control Plane provides the shared language for managing it.
Build against infrastructure.
Horizon is built to make the control plane the interface. Configure workloads, inspect deployments, and integrate through clear HTTP APIs.
Read the developer docsGET /api/models
POST /api/deployments
{
"modelId": "...",
"workerId": "..."
}Your machines do not need to look alike.
The current MVP focuses on trusted, team-controlled hardware. Support is shaped around what the worker can report and what the runtime can use.
Infrastructure is also about what stays separate.
Workers do not touch PostgreSQL
Worker communication goes through the Control Plane. Database access remains a control-plane concern.
Structured commands
The system is designed around explicit deployment and runtime messages rather than arbitrary shell access.
Metadata stays distinct
Model metadata and configuration are separate from model weights and local runtime state.
A practical starting point for local infrastructure.
Use available hardware for local model serving and development workflows.
Coordinate compatible machines across a research environment.
Pool trusted hardware without immediately renting more cloud capacity.
Expose internal compute through one operational control plane.
Keep the boundaries clear.
Workers communicate with the Control Plane. PostgreSQL remains behind it. The deployment path is explicit from request to runtime.
Your machines are compute.
Connect compatible hardware, deploy models, and build on top of a distributed AI runtime.