Local, cloud, and open models

Home Learn Local, cloud, and open models

In short

A local AI model runs on your computer; a cloud service runs it on remote computers. Where it runs and what its license allows are separate questions. Available model weights don’t automatically make a system open source or grant permission for every use.

Why it matters

You may care about offline access, who receives your information, how the system is maintained, or whether you can adapt it. Those concerns won’t be answered by “local” or “open” alone. A local application can still use remote features. A hosted service can provide strong controls, but you need to understand the whole setup and the terms that apply to your use.

How it works

Three questions, not one

First, ask where computation happens. A device can perform inference locally if it has the needed software and hardware. Inference is the process of producing an output from a trained model. A cloud service runs computation on remote infrastructure. NIST’s cloud definition describes shared computing resources delivered through network access.

Second, ask what components are available. Weights are learned numerical settings. Releasing them can let other people run or adapt a model, but doesn’t necessarily provide the training code or information about the training data. You need to know what the release includes.

Third, read the license. A license sets conditions for using, modifying, or sharing the supplied material. The Open Source Initiative’s AI definition addresses freedoms and the preferred form for making modifications, including data information, code, and parameters. “Open weights” and “open source” can therefore describe different levels of access. A label on a download page isn’t a substitute for its terms.

Compare the complete operating setup

Look at the application around the model. Does it send prompts, usage records, or error reports over the network? Can an extension call another service? Who can reach the local service, and who installs updates? Local processing addresses one part of the data path.

Consider the work of maintaining the system as well as its output. Someone needs to manage software updates, model provenance, access, logs, and backups. I’d compare a narrow task across the complete setups, with the same source and review criteria, before deciding that a deployment location solves every concern.

Example

Illustrative example: this scenario is invented.

Eli works in IT support and security. They compare a local application and a hosted assistant using a help article. The local option can generate a response on the device, but an enabled extension also offers web retrieval.

Eli records that distinction instead of assuming the whole application is offline. They examine the model’s license separately from the application’s settings. The useful question is whether each setup meets the requirements for the intended job, including who maintains it and where information travels.

What it doesn’t mean

  • Local processing doesn’t guarantee privacy. Logs, connections, shared devices, and access still matter.
  • Downloadable weights don’t grant unrestricted rights. Read the license and release information.
  • Cloud doesn’t describe a single security standard. Evaluate the service and its configuration.

Try it

Spend about ten minutes asking an assistant to make a comparison checklist for two imaginary setups. One runs on a laptop but has optional web search. The other runs remotely and is managed by a small team.

Have it separate computation location, network traffic, license, access, and maintenance. Ask what evidence would answer each question, without recommending a product or inventing terms. Check that the checklist keeps deployment location separate from openness and security.

Keep going

Sources

Share this page