Deployment

Turn a trained model version into an online inference service for your application.

A deployment connects a model version to API calls. Completing training does not automatically make a serving endpoint ready: loading, health checks, and routing must also succeed.

Choose a deployment method

The current deployment page exposes these options. Availability depends on model compatibility and configured resources.

Method	Requirements
LoRA · own GPU cluster	The base model is supported by the cluster and matches the adapter
Tinker · platform	The version has an inference checkpoint; the platform resolves its base model

Start from a version in My models (我的模型) or choose one on Deploy (部署). Confirm that the base and artifact match. See Pricing for dedicated GPU inquiries.

Deployment list and readiness checks · Screenshot placeholder · final product capture to follow

Verify readiness

Check	Meaning
Loading	Model or adapter has loaded
Health	The inference service passes its health check
Routing	The calling endpoint can route to the service

A deployment record alone does not establish availability. Review state and errors if a readiness check fails. A stopped service needs to become available again before requests can succeed.

Connect your application

Open Model calls → API access (API 接入), choose the model and API key, and use the generated configuration. The current Alpha environment exposes an OpenAI-compatible HTTP interface with /v1/chat/completions for chat requests.

In the example below, set MINT_BASE_URL to the base address shown on the access page, MINT_MODEL_ID to that service's Model ID, and MINT_API_KEY to your key.

# Set MINT_BASE_URL, MINT_API_KEY and MINT_MODEL_ID from API access.
curl "${MINT_BASE_URL%/}/v1/chat/completions" \
  -H "Authorization: Bearer $MINT_API_KEY" \
  -H "Content-Type: application/json" \
  --data-binary @- <<EOF
{
  "model": "$MINT_MODEL_ID",
  "messages": [{"role": "user", "content": "Hello"}],
  "max_completion_tokens": 128
}
EOF

Use the API access Model ID

Training version IDs, checkpoint references, and API Model IDs identify different resources. Use the identifier supplied by the access page rather than substituting a version number or sampler_path.

Verify one request before integrating application traffic. If the model is unavailable or calls fail, see Troubleshooting.