Open-Weight Models vs Hosted AI Services
Understand the difference between hosted, closed models and open-weight models, and when each makes sense.
There are two very different ways to get a model into your app, and the choice shapes your costs, your privacy and what you’re allowed to do.
A hosted model is like ordering a dish in a restaurant. The provider runs the model on its own GPUs and you send it requests through its API. You never see the weights. Getting started takes minutes, you pay only for what you use, and the strongest frontier models are often available only this way. The trade-offs: your data travels to the provider, prices and models can change under you, and you depend on their service staying up.
An open-weight model is like being handed the recipe. The maker publishes the weights, the file of numbers itself, for anyone to download, and Meta’s Llama, Mistral’s models, Qwen and DeepSeek are well-known examples. Now you choose who does the cooking. You can run it on your own hardware, so data never leaves your building and it can even work offline. You can fine-tune it on your own examples (LoRA, from Track 2). Or you can pay a hosting provider to run your chosen open model for you.
Freedom comes with chores. Running models yourself means GPUs, serving engines, monitoring and updates: all of Track 4 becomes your problem. And the price sums flip. A hosted API charges per token, so it’s cheap at low volume. Your own GPU costs the same every hour whether busy or idle, so it only pays off once you have enough steady traffic to keep it busy.
One more detail: “open weights” isn’t quite “open source”. You get the trained numbers, usually not the training data or full recipe, and the licence may limit how you can use the model, so it’s worth reading before you build on one.
Key takeaways
- Hosted models: the provider runs them, you pay per token, often the strongest models, but your data leaves your hands.
- Open-weight models: the weights are published, so you can run, fine-tune and keep data in-house.
- You can also have an open-weight model hosted for you: open model, someone else’s GPUs.
- Self-hosting has fixed costs and engineering work; it pays off only with enough steady traffic.
Check your understanding
Go deeper (more technical)
The Open Source Initiative published a definition of “open-source AI” in 2024 that requires enough information about the training data to recreate a substantially similar system. Most popular “open” models release weights but not the data, which is why many people prefer the term open-weight.
Hybrid setups are common: a hosted frontier model for the hardest requests, a cheaper open model for the bulk, or a hosted provider’s private-deployment option that keeps data inside your own cloud account.
The simulation’s prices are illustrative.