Introducing Epilude Model 4: our best on-device dictation model yet

For most of its history, dictation has meant your voice traveling to someone else's computer. We built Local Mode to end that trade-off, and Model 4 is our strongest argument yet that you do not have to choose between quality and privacy.

Today we are shipping Epilude Model 4, the fourth generation of the on-device model that powers Local Mode. It is the largest quality jump since Local Mode launched: the same 1.5 GB download and the same speed, but a measurably better writer, with nothing you say ever leaving your Mac.

What Model 4 is

Local Mode runs two models directly on your Mac. A speech model turns your voice into text, and a cleanup model turns that raw transcript into writing you would actually send: it removes filler, fixes punctuation, resolves the false starts and mid-sentence corrections that are natural in speech, and matches the tone you asked for.

Model 4 is a new generation of the cleanup model: our own fine-tune, built on the Qwen family of models and trained specifically for dictation cleanup. The speech model is unchanged in this release. Together they fit in the same footprint as before and run at the same speed on the same hardware.

The only errors that matter are the worst ones

Average quality is the wrong way to measure a dictation model. A model that writes beautifully 95 percent of the time but occasionally turns "let's meet Thursday" into "let's meet Monday" is worse than useless, because you stop trusting everything it writes.

So we grade our models on a zero-tolerance error class we call critical errors: any output that changes the meaning of what you said, or that leaks the model's instructions into your text. Our shipping rule is absolute. A new model may not introduce a single new critical error compared to the model it replaces, no matter how much better it is on average.

How we evaluate

We maintain an internal benchmark of 90 dictation scenarios spanning punctuation, formatting, self-corrections, tone control, multiple languages, and mixed-language speech. Every candidate model runs the full suite repeatedly, and every output is graded by frontier-model judges casting multiple independent votes. A scenario only counts as passing if it passes in every repeated run, and we re-grade identical outputs across runs so that grader noise never gets mistaken for a model change.

We do not publish this benchmark, and results on it are not comparable to anything external. But it is how we hold ourselves accountable, and it is the bar every Epilude model has to clear before it ships.

Results

On that internal benchmark, Model 4 passes nine more scenarios than Model 3, and critical meaning errors are down 56 percent:

A scenario counts as passed only when it passes in every repeated judged run.

The improvements concentrate exactly where dictation models tend to embarrass themselves: mid-sentence self-corrections, dates and weekdays, mixed-language speech, and casual tone that should keep your hedges instead of scrubbing them into corporate flatness. It is easiest to show:

Earlier generations would sometimes keep both days in that first sentence, or the wrong name in the second. The third stays in the languages you spoke instead of being silently translated. And the last one, cleaned in a casual style, keeps "kinda" because you said it and meant it.

What we learned building it

Two findings from this release will shape every future model we train.

First, label quality beats data volume. When we audited the hardest examples in our training data, we found that a surprising share of them were teaching the model the wrong lesson, in ways invisible to automated filters. Fixing those examples, rather than adding more data, is what unlocked the self-correction improvements above.

Second, we shipped an average, on purpose. We trained several otherwise-identical runs of Model 4 that differed only in their random seed, and found that each run had its own small set of rare failure modes. The model we shipped is the weight average of those runs. It outperformed every individual run and carried none of their individual quirks. Weight averaging is a known technique in the research literature; what surprised us is how decisively it beat picking a winner.

Limitations

Model 4 is not perfect, and we would rather tell you where it struggles than let you find out. Long, rambling passages remain the hardest class of input: on our benchmark, the remaining failures concentrate there, especially long spoken questions. That is the focus of the next generation. Cloud mode is still ahead on our internal suite overall, and Local Mode remains the choice when privacy comes first.

Privacy, requirements, availability

Local Mode processes everything on your Mac. Your audio and text are never sent to us or to anyone else. It is included with Epilude at no extra cost.

Model 4 requires Apple Silicon. Full AI cleanup runs on Macs with 16 GB of memory or more; on smaller machines, Local Mode does on-device transcription with basic formatting, and still never sends anything off your Mac. If you already use Local Mode, the app will prompt you to download the updated model, about 1.6 GB, after you update. New to Local Mode? Open Settings → On-device models and switch it on.

What's next

Long-input quality, deeper multilingual coverage, and the same evaluation discipline applied to more of the product. If Model 4 changes how dictation feels on your Mac, or if you catch it making a mistake, we want to hear about it through the Help Center.

The Model 4 cleanup model is built on the Qwen family of models. Local Mode's speech recognition is built on optimized open models. Full attributions are in the app under Acknowledgements.