Every autocomplete tool has to run a language model somewhere. That "somewhere" is the entire question. Run it on a server, and your keystrokes have to travel there first. Run it on your own machine, and they never leave. The suggestion you see on screen can look identical either way — the difference is invisible until you ask where the computation actually happened.
This isn't an abstract distinction. It determines what a company can technically do with your text, what a breach of their servers exposes, and whether your autocomplete tool works on a plane with the Wi-Fi off. Here's what's actually different under the hood.
How Cloud Autocomplete Works
A cloud-based writing assistant is typically a lightweight client — a browser extension or a small background process — paired with a large model running on a remote server. The flow looks like this:
- You type a few words in a text field.
- The client captures the surrounding text and sends it over the network to the vendor's API.
- A large model on their infrastructure generates a completion.
- The suggestion is sent back and rendered as ghost text in your field.
This round trip happens on every keystroke pause, which means a continuous stream of what you're writing — draft emails, chat messages, internal notes — passes through a third party's servers. Some vendors log this traffic for model improvement, abuse detection, or debugging. Some don't. From the outside, as a user, you generally cannot verify which is true, and the terms can change with an update to the privacy policy.
How a Local LLM Works
An on-device tool inverts the architecture. The model itself — not just the client — lives on your machine.
The model is downloaded once
A compact, quantized model file is downloaded to disk during setup. After that, no further network access is required for it to function.
Inference happens on your CPU/GPU
When you pause typing, the model runs locally through an inference engine, reading the nearby text directly from memory — it never leaves the process.
No network round trip
There's no server to send text to, so there's no request to intercept, log, or retain. The suggestion is computed and discarded in the same local process.
Typeahead for Windows is built this way. It runs a local model — Qwen 3, served through llama.cpp — directly on your PC. The only network activity it needs is downloading the model file and checking for app updates; generating suggestions requires no connection at all, and your typed text is never transmitted anywhere.
What "No Telemetry" Actually Means
"Privacy-friendly" is a phrase that gets attached to a lot of products with very different implementations underneath. It's worth being specific about what a genuinely local architecture rules out, technically, not just contractually:
With no server involved, there's no text-logging pipeline to misconfigure, no analytics dashboard that could accidentally include message contents, and no dataset of user text that could later be involved in a breach, an acquisition, or a change in data-use policy. The absence of a cloud path isn't a policy promise — it's a property of the architecture.
The Trade-Off: Model Size and Capability
None of this is free. A model that runs entirely inside 400 MB–2.5 GB of disk space, on a laptop CPU, is smaller than the massive models cloud services run on datacenter GPUs. For open-ended tasks like drafting a full essay from scratch, a giant cloud model still has an edge in raw creative range.
But next-phrase autocomplete is a narrower task than open-ended generation — it's predicting a plausible short continuation of a sentence you've already started, not composing new content from nothing. That's a task a small, well-tuned local model handles well, which is why the size trade-off is much smaller for this use case than for general-purpose chat.
| Property | Local LLM (Typeahead) | Cloud autocomplete |
|---|---|---|
| Where inference runs | Your device | Vendor's server |
| Text sent over network | Never | On every suggestion |
| Works offline | Yes | No |
| Account required | No | Usually |
| Exposure if vendor is breached | None — no text stored server-side | Depends on retention policy |
Where This Matters Most
The gap between local and cloud isn't just theoretical for anyone typing sensitive material — support agents handling account details, healthcare or legal staff drafting notes, or anyone writing under a confidentiality agreement. It's also relevant if you simply write on a train, in a basement office, or anywhere connectivity is unreliable: a cloud tool goes silent, a local one doesn't notice.
Typeahead treats this as a default, not a setting. Password fields are always ignored, and password manager and finance apps are protected out of the box, with no toggle to turn that protection off — on top of the fact that typed text was never leaving the device in the first place.
Frequently Asked Questions
Does Typeahead ever send my typed text anywhere?
Can I verify a tool is actually local and not just claiming to be?
Is a small local model as good as a large cloud model?
Does Typeahead require an account or sign-in?
What does Typeahead download from the internet, if anything?
Get Typeahead
Suggestions generated on your device, not on a server. One-time $9.99, 7-day free trial.
Get it on Microsoft Store