Three Ways to Get a Model
Compare hosted APIs, open-weight models, and managed services on control, effort, quality, and exit.
Objective
By the end of this lesson, you can explain the trade-offs of hosted APIs, open-weight models, and managed cloud services, and name which parts of a system a team should own no matter which it picks.
The three options
Almost no community organization trains a large language model from scratch. Training a frontier model takes data, hardware, and staff on a scale few institutions have. So the real choice is how you get access to a model someone else trained.
Hosted API. You send text to a provider's servers and get text back. The provider runs the model. You pay nothing to host it, and you get strong models quickly. You also send your data to someone else and depend on their uptime, terms, and model retirement schedule.
Open-weight model. The model's trained numbers (its weights) are published, so you can download and run it yourself. Llama, Mistral, Gemma, and Qwen model families are examples, each with its own license. You control where data goes. You also take on hardware, updates, security patches, and the skill to run it.
Managed service. A cloud provider runs models for you inside your cloud account, often with business contracts about data handling and region. It sits between the two: less work than self-hosting, more control than a public API, and more setup than either.
The trade-offs side by side
| Factor | Hosted API | Open-weight, self-hosted | Managed cloud service |
|---|---|---|---|
| Time to first working version | Fastest | Slowest | Medium |
| Where data goes | Provider's systems, under their terms | Your systems only | Your cloud account, under a contract |
| Staff skill needed | Low to medium | High | Medium |
| Model quality ceiling | Highest available | Varies; often a step behind on hard tasks | Depends on which models are offered |
| Exit difficulty | Depends on how much you own | Easiest | Medium |
| Who patches and monitors | Provider | You | Shared |
Read the license and the data terms for every option. "Open" does not always mean you may use a model for any purpose. "We don't train on your data" may have exceptions, such as logs kept for abuse monitoring. Ask in writing.
What you own in every case
The model is the part you are most likely to swap. The parts that hold your judgment are the parts you should own:
- Your prompts, versioned like code (Module 3).
- Your retrieval layer: documents, chunking, and index (Module 4).
- Your test sets and evaluation results (Module 7).
- Your guardrails and permission rules (Modules 7 and 8).
- Your logs and metrics (Module 10).
This is the blend pattern. You rent the model and own everything that makes the system yours. If a provider changes terms or retires a model, you run your test set against a replacement and switch with evidence.
Quiz: Options check
A national example
Practice
What would you own?
Go deeper: Own the code: building internal tools over rented software
Discussion
Ask. Answer. Build together.
Loading…