Why the first launch is slow, and what the Neural Engine is doing

13 August 2026 · 3 min read · Muteeb Mehraj

The first time you dictate with Bellow there is a wait. After that there is not. That gap confuses people, so here is exactly what is happening in it.

The pause

Bellow bundles Whisper Large-v3 Turbo, roughly 1.5 GB of model. It is included in the download rather than fetched later, which is why the disk image is large and why the app works with the network off from the very first launch.

A bundled model still is not a ready model. Before Apple's Neural Engine can run it, the weights have to be compiled into a form the hardware accepts. That compilation happens once per machine, on first use, and it is not fast.

Measured on a Mac mini with the compile cache cleared, so it genuinely started from nothing:

• 66.90 seconds on one run.
• About 30 seconds on another.

The variance is normal; it depends on what else the machine is doing. On slower or busier Macs it can take longer.

Then the result is cached, and every launch afterwards loads in a few seconds.

Why this was a bug even though nothing was broken

For a while, this worked correctly and looked broken.

You would finish setup, land on the dashboard, press your trigger key, and nothing would happen. No error, because there was no error. The model was compiling. The app simply had no way to say so.

That is a bad first thirty seconds for software whose entire pitch is that it responds instantly. The person cannot distinguish "working hard" from "not working," and they have no reason to give it the benefit of the doubt on first run.

Bellow now shows a banner while it compiles, with the elapsed time counting up, and confirms when it is ready. It appears only on first run; on later launches the model is already compiled, so the banner never shows at all.

The code did not change. The honesty did.

What the Neural Engine buys you

The compile is worth waiting through once.

The Neural Engine is dedicated silicon for the matrix operations that dominate model inference. Running Whisper there rather than on the CPU means transcription finishes in a fraction of the audio's duration, and it does it without spinning fans or draining battery the way sustained CPU inference would.

It is also why Bellow requires Apple Silicon. There is no Intel build, and not because we did not get to it: an Intel Mac has no Neural Engine, so the model would run on CPU. It would work, slowly and hotly, and it would feel like a different and worse product. We would rather not ship that than ship it and have it define what people think Bellow is.

Turbo, and the trade we took

Large-v3 Turbo is a distilled variant of Whisper Large-v3. It reaches first token roughly four times faster than full Large-v3, at a small accuracy cost on difficult audio.

For dictation that trade is clearly right. You are speaking clearly, close to a microphone, in a quiet-ish room, and you are waiting for the text before you can continue. Latency is the thing you feel on every single use. The accuracy difference shows up mostly on the kind of audio dictation does not produce: crosstalk, distance, heavy background noise.

If it seems stuck

Give it a few minutes on first run. The banner shows elapsed time so you can see it is progressing rather than hung.

If you want to reproduce a genuinely cold start, for example to check what a new user experiences, clear the compile cache:

rm -rf ~/Library/Caches/com.apple.e5rt.e5bundlecache

That cache is shared with other Neural Engine apps on your Mac. Deleting it is safe, since the OS rebuilds it on demand, but the next launch of any of those apps will also be slower than usual, once.