Package a model with Truss, deploy it to Baseten, call its predict endpoint, and promote it to production

Documented steps

1. Install the Truss CLI: pip install truss (or uv tool install truss).
2. Authenticate: truss login --browser. This stores your Baseten API key locally; no manual key entry needed.
3. Create the model directory and a config.yaml declaring model_name, resources.accelerator (for example L4, A100, H100), and either a python model.py implementation or a framework-specific block such as a TensorRT-LLM checkpoint repository.
4. Deploy a development deployment: run truss push from the project directory. The CLI prints the model id (an 8-character handle) and a development deployment URL.
5. To publish straight to production instead, run truss push --promote, which creates the deployment and immediately makes it the production deployment.
6. Recover the model id later via the management API: GET https://api.baseten.co/v1/models with header 'Authorization: Bearer <BASETEN_API_KEY>'. The response array carries id, name, production_deployment_id and development_deployment_id.
7. Call the model: POST https://model-{model_id}.api.baseten.co/environments/production/predict with the same Authorization header and a JSON body matching your model's input schema. Use /environments/development/predict or /deployment/{deployment_id}/predict to target other deployments.
8. For OpenAI-compatible Trusses, point the OpenAI SDK at base_url https://model-{model_id}.api.baseten.co/environments/production/sync/v1 with the Baseten key as api_key.
9. Watch build and deployment logs in the Baseten console; a deployment stays in a building state until the image is ready, and predict calls against it will fail until it is active.
10. Official documentation: https://docs.baseten.co/development/model/build-your-first-model | https://docs.baseten.co/inference/calling-your-model | https://docs.baseten.co/reference/management-api/models/gets-all-models