0.4. Providers
In one glance
- You will: Distinguish model hosting, open-source software, open weights, and offline test doubles.
- You need: Nothing beyond a browser.
- Time: about 10 minutes, concept.
Which model path should you choose?
Use Gemini through an API key for the main course path.
| Path | Requirements | What it proves |
|---|---|---|
| Gemini API — default | Account, API key, network, and available quota | Behavior of the selected hosted model |
| Ollama/Qwen3 — optional | Model download and suitable local hardware | Behavior of that local open-weight model |
| Offline workshop | Locked Python runtime | Tools, state, approvals, and grader behavior |
| Recorded answers | Course fixtures | How to inspect and grade known outcomes |
Hosted Gemini avoids requiring GPU hardware. Free-tier eligibility and quotas vary, so the course provides useful offline work when access is unavailable. The optional cloud platform uses Vertex AI and GKE with separate authentication and billing.
What is open source in this course?
The course uses open-source agent and platform software, while the default model service is proprietary.
ADK, agentgateway, kagent, MLflow, OpenTelemetry, Prometheus, Grafana, and Ollama have their own open-source licenses. The course's text is CC BY 4.0 and code is MIT. Qwen3's selected artifact has Apache-2.0 open weights; that does not claim every training artifact is open.
Gemini, Vertex AI, GKE, and other Google Cloud services are proprietary hosted services. The optional local path permits completing model-backed labs without a hosted model account, subject to hardware and measured model capability.
What should you record when comparing models?
Record the exact model, prompt source, dataset, transport, date, sampling configuration, and observed outcomes.
A model replacement changes behavior even when the API shape remains compatible. Compare critical tool, approval, and evidence cases, then task success, latency, and token usage. Never reuse a previous model's evaluation result as proof of a replacement.
The course keeps a compatibility pin shared by the default Gemini and gateway profiles. A newer release is a candidate for evaluation, not an automatic upgrade.
What proves this page worked?
Choose your main path and identify its access requirement.
You are done when:
- You can distinguish open-source software, open model weights, and a hosted API.
- You know that free course access does not promise unlimited free inference.
- You know which offline exercises remain available when model access is unavailable.
Continue to 1.4. Providers when you need to configure the chosen path.