A downloaded model running on your own computer can avoid a hosted provider's per-token usage bill. That does not give it an unlimited context window, an unlimited answer, or unlimited computing capacity. The useful distinction is between how often you may use it and how much text it can process in one request.
This matters if you are choosing local AI to avoid message caps, write frequently, or work offline. It is less helpful if your real need is to read a very large document in one pass: that is a context and memory problem, not a billing problem.
Four limits that are easy to confuse
| Limit | What it means | Does running locally remove it? |
|---|---|---|
| Hosted usage allowance | Messages or tokens included by a service | A fully local model need not use that service, but a selected cloud provider still does |
| Context length | Instructions, conversation, supplied text, and answer space for a request | No; the model, app, and available memory constrain it |
| Output allowance | How long one answer may be generated | No; the app or model runner may cap or stop the answer |
| Computing resources | Memory, processing time, disk space, power, and heat | No; your computer supplies these |
A token is a unit of text used by a model. It might be a word fragment, punctuation, or another piece of text. Do not treat “one token equals one word” as an exact conversion, especially for code or non-English text.
Check that the request is really local
- Install a model runner and download a supported model.
- Select that local model in the chat app. A local interface can still call a hosted provider.
- Complete a short conversation, then test an offline restart.
- Check optional search, document, and tool providers separately.
If the same new chat works after restarting offline, you have a practical check that this particular chat does not need a hosted model for that request. It does not prove every feature, later upgrade, or connected tool has the same behaviour.
Why the model stops halfway through an answer
An answer may hit its output allowance, reach a stop sequence, exhaust usable context, or fail because the computer runs out of memory. Those causes require different fixes.
Ollama's Modelfile reference separates num_ctx, which controls context, from num_predict, which controls generation length. Other apps use their own labels for these settings. Inspect the setting for the actual model and app rather than copying a maximum from another product.
If a draft cuts off, ask for the remaining section with enough source context, or increase the output allowance when the app and memory permit it. If the model loops or repeats, stop the request and use answer troubleshooting. An unrestricted generation setting can waste time without improving the answer.
Why a large document still does not fit
Use the token counter to estimate the supplied text. Allow for your instructions, the existing conversation, document passages, tool output, and the answer. Use the model's own tokenizer for an exact count; another tokenizer gives a planning estimate.
Then check context windows and the memory calculator. A published long-context limit does not mean your computer can hold it. Document search may supply selected relevant passages instead of the entire document; that can be useful, but it changes what the model sees.
For example, ten separate short chats may be affordable to run repeatedly on a laptop. A single chat containing ten long documents may fail even though the combined number of tokens over the day is the same.
Costs and terms still apply
An already-owned computer can have no new purchase cost while still using electricity, battery life, storage, and your time. Optional cloud models, paid app features, or remote tools can introduce their own charges. The app's license and the downloaded model's license are separate too.
Use the cost calculator with your intended workload. Its estimate can compare money costs under supplied assumptions; it cannot promise equal quality or a fixed number of useful answers per day.
The practical goal is a local model you can use repeatedly for your actual task, with manageable waits and enough memory. Calling it unlimited conceals the limits that decide whether it will be useful.