Client App
Run your containers on GPUs worldwide.
Publish a container once, then send tasks to the right GPUs anywhere. Training, rendering and batch jobs, with live logs and pay-per-task pricing.
How it works
Three steps from container to result.
- 01
Publish a source
Package your code, dependencies and execution parameters as a container. A source is a reusable definition you can run again and again.
- 02
Submit tasks
Send tasks that reference your source, with inputs and requirements: GPU type, VRAM, budget, runtime limits and priority.
- 03
Get results
The network routes each task to a matching GPU, runs it in isolation, verifies the output and returns it, with full logs and run metadata.
Features
Everything you need to run GPU jobs.
Sources & templates
Start from templates for popular models and frameworks, or bring any container.
Routing preferences
Set hardware type, performance and priority rules. The router does the rest.
Live logs
Follow every task in real time, from routing to result.
Spend per source
See exactly what each workload costs. Pay only for completed tasks.
Automatic failover
If a node drops, the task resumes on another GPU from the last checkpoint.
Full task history
Every run is recorded with inputs, outputs, timing and the hardware it ran on.
Client App or Relay?
Bring your own code, or call a model.
Use the Client App when you run your own containers. If you just need a model behind an API, Relay is faster to start with.
Good fit
- Fine-tuning and training
- Rendering and video pipelines
- Batch and research jobs
- Custom models and checkpoints
Works with
- Docker containers
- PyTorch, TensorFlow, JAX
- Python SDK and REST API
- Webhook callbacks
Submit your first task.
Open the Client App, publish a source and run it on the network.