Deployment
Choose a service
| Deployment goal | Entrypoint | How to check it |
|---|---|---|
| Parse interactively on one machine | mineru-kit webui manages a local parsing worker by default. |
Open the address printed at startup and convert a file using the WebUI guide. |
| Offer document parsing to applications | mineru-kit api-server exposes the asynchronous Tasks API. |
Inspect /v1/health and /v1/tiers, then run the single-file HTTP example. |
| Put multiple GPUs or existing parsing services behind one address | Deploy a Router. | Submit the same Tasks example to the Router's address. |
| Return an attachment after a single-file upload | Run file-parse in front of a Tasks service. | Check the attachment or the status URL in a 202 response. |
| Provide remote VLM inference | Run mineru-kit vlm-server or use an existing OpenAI-compatible model service. |
Set model.vlm.server_url on the parser and try a Standard or Advanced parse. Remote service configuration |
Prepare hardware and models
- Use
mineru-kit hardware info --jsonto inspect devices, runtimes, and existing calibration records. Usemineru-kit hardware detectfor a tier recommendation. Initial ONNX calibration may download a model. Tiers and runtimes - The base package can parse native documents directly. For model inference, select the optional modules or an NVIDIA Docker image for the target device. For multi-GPU assignments, see GPU and engine parameters.
- Download and verify the models for the backend and engine you will run.
mineru-kit models showreports selection and readiness. Use the exit code frommodels verifyto check required files. Complete both steps before going offline. - Start the service and inspect a result. The server's
--tiersets the ceiling for PDFs and images. Native documents use Flash. Use native HTML to check the request and download path, then parse a representative PDF or image at the inference tier you need.