Public guide
Which model fits?
A local model and a hosted frontier model do different jobs. Use this page to match the model to the task. It explains the choice. It does not rank products or publish test scores.
On hardware you operate
When a local model makes sense
A local model runs on machines you control. For a client, that is hardware in their own environment. The choice changes where the prompt lives, whether the work needs a provider online, and who operates the system.
- Privacy, with limits
- The prompt can stay on hardware you control. That matters when the material should not be sent to a model provider. Privacy still depends on the machine, its logs and backups, who can sign in, and which tools the workflow calls.
- Offline capability
- After the model is installed, it can answer without a live connection to a provider. First-time setup, model updates, and any tool that calls the network still need connectivity.
- Control
- You choose the model and when it changes. You also take on evaluation, updates, and the day-to-day operating routine.
- Hardware and operating cost
- The cost shows up as machines, power, and the people who run them. A local model still has a bill. Whether that bill makes sense depends on the work and on hardware you already have. This guide leaves prices out.
A provider's service
When a hosted frontier model makes sense
Frontier models such as GPT, Claude, Gemini, and Grok run in a provider's cloud. They are a strong option for many hard, open tasks. You rely on that provider, and the cost follows how you use the service.
- Difficult reasoning and range
- Hosted frontier models are often stronger at difficult reasoning and at broad, open-ended work. Treat that as a tendency. Check the task you care about, because a general reputation can miss your case.
- Provider dependence
- The provider hosts the model, sets the terms, and can change the model, the limits, or the service. Your workflow follows those changes unless you have another path.
- Usage cost
- Hosted use is commonly billed by token, seat, or subscription. Starting is often simpler than standing up hardware. What you pay at your own volume is a separate question. This guide leaves rates out.
A practical comparison
Which fits your task?
Read each row as a cue for the job in front of you. These are not measurements, prices, or a ranking of products.
| Task | Local model | Hosted frontier model |
|---|---|---|
| Material that should stay in your environment | Fits when the machine, logs, and access match that requirement. | The prompt and attached context go to the provider under their terms. |
| Work that must continue without that provider | Can answer once the model is installed, including during a provider outage. | Available while the provider and the path to them are up. |
| Difficult reasoning or an open-ended problem | Worth a trial on the actual task. Some local models cover a bounded step. Broader problems are where they more often fall short. | Often the stronger place to start for hard reasoning and broad problems. Confirm it on your task. |
| A general assistant with a wide brief | The range is the model you install and the tools you connect. | Usually the wider model and toolset, inside that provider's product. |
| The same private step, repeated | You can pin a version and keep the step in your environment. | Simple to start. Cost follows volume, and the provider can change the model under you. |
| A first result before hardware is ready | Needs a machine, setup, and someone to operate it. | Usually the faster way to try a task. You accept their terms, limits, and pricing. |
Local and hosted together
Use both when the work splits
A useful system can send one step to a hosted model and another to a local model. The knowledge stays with the client. The model follows the step.
-
01
Split by sensitivity
Explore a hard question with a frontier model, using only material you are allowed to send. Run the private, repeating step on a local model.
-
02
Reduce before you send
Keep sensitive records on the local model. When a frontier model is still useful, send a version that leaves those records out.
-
03
Design in one place, run in another
Shape the workflow with a frontier model, then pin a local model for the version the team operates.
Quality, access, and cost still need a check. Using both is a way to give each step a fitting model.