Why a Gemma model cleans up your text

Didn't macOS already have a model for this?

It does, and cleanup started there. An earlier version handed dictations to Apple's on-device model, the one behind Apple Intelligence, and fell back to a set of fixed rules for everything else.

"Everything else" turned out to be a lot. Apple's model only runs when Apple Intelligence is turned on, has finished downloading, and supports the language you are speaking. A Mac with it switched off, or a dictation in a language outside its list, got the rules: fillers and spacing tidied, nothing smarter. Two people with the same app could get two quite different results from the same sentence.

Why Gemma, then?

Three things had to be true at once. The model had to run on this Mac, so that dictation keeps working with the network off and your text never leaves the machine. It had to handle the languages people actually dictate in, including several in one sentence. And it had to be small enough to sit in memory beside a speech model in a background app.

Gemma 4 meets all three. It covers 140+ languages, so a dictation in Ukrainian or Thai gets the same pass as one in English. Its weights are published for anyone to run locally, so nothing about using it requires a call to Google. And it comes in two sizes, so there is a choice that does not involve changing model families.

The rule-based fallback went away with the move. If the model is not there, you get the recognizer's words unchanged, which is easier to reason about than a third, weaker kind of cleanup.

Why such a small model?

Because the job is copy-editing, not writing. The default, Gemma 4 E2B, is small enough to download without a second thought and fast enough that cleanup keeps up with dictation on every Mac Dictera runs on. E4B, the larger one, is roughly twice the effective size. It is noticeably better at translation and at following a spoken instruction, and barely better at removing an "um", which E2B already does well. You can switch between them in Settings ▸ Engine Models or from Quick Controls, and keep both downloaded; only the one in use is loaded into memory.

Both are quantization-aware builds: Google trained them to run at 4-bit rather than rounding a full-precision model down after the fact. For the same download size, that holds up better than the usual compressed copy.

What is it allowed to change?

A short list, and nothing outside it. It removes fillers, false starts and words said twice. It fixes punctuation, capitalization and spacing by the rules of the language you spoke. It fixes a clear grammatical error with the smallest change that works, an unambiguous misspelling, a word the recognizer wrote in the wrong script, and a brand name where the intended one is obvious.

It does not rephrase, summarize, translate, change your tone or swap a correct word for a synonym. When it is unsure, it keeps what you said. It runs with greedy decoding, so the same sentence gets the same edit every time. Two models per dictation walks through the full instructions, and the check that throws away an edit no copy-editor would have made.

A bigger model would not change that list. The list is what makes cleanup safe on text you are about to send in your own name, and a small model that keeps to it is a better fit than a large one that could do far more.

How do you know the model you have is the one we tested?

Each model is pinned to one exact published version of its weights, and the file's checksum is verified before it is installed. Two people on the same build of Dictera run identical weights, and a download that was interrupted or altered in transit is rejected rather than loaded. The download comes from Hugging Face, which is on the list of hosts the app contacts, and it happens once.

When a newer revision of the same model ships, Settings shows it as an update with a line saying what changed. Nothing swaps underneath you.

Does the same model do anything else?

Yes. The Gemma you pick also runs Translation, Correct Selection and, unless you point it elsewhere, the AI Assistant. Each feature gets its own instructions; the weights in memory are shared, so a second feature does not cost a second multi-gigabyte download.

Any of them can instead run on a model you host yourself. Pointing cleanup at Ollama or LM Studio keeps the same instructions and the same safety check, with your own model behind them. Gemma stays the default because it needs nothing from you beyond one download.