Twice as fast, at half the size

Twice as fast, at half the size. Glimpse 1.3.1 rebuilds how Parakeet TDT V3 runs on Apple silicon. Short dictations finish in under 30 milliseconds, a 9 minute recording takes under 2 seconds, and the download drops to 549 MB. Every word comes out just as accurately as before, entirely on your device.

Remarkably fast

On an M2 Pro, compared with Glimpse 1.3.0:

• Short dictations, 2.5 times faster. 27 ms instead of 68 ms. Clips under 5 seconds take just 16 ms.
• Long recordings, 2.4 times faster. A 9 minute recording takes 1.8 seconds instead of 4.2.
• Half the download. 549 MB instead of 1.1 GB.
• Lighter on memory. About a third less while you dictate.

Engineered end to end

The speedup comes from every stage of the pipeline. The decoder now reads the model's 8-bit weights directly, with a quarter of the memory traffic. The Neural Engine encodes the next part of a recording while the current one is decoded. And the new encoder carries 15, 10 and 5 second versions, so a short dictation runs on the smallest one that fits.

Just as accurate

Every change was verified against 14,058 recordings: the full LibriSpeech test sets, FLEURS in all 25 languages Parakeet supports, long recordings, silence, numbers and names. No language got worse, and words from your dictionary come out right as often as before.

Faster on Windows, too

The same work speeds up the local models on Windows, and brings Nemotron to the graphics card on both platforms.

• Parakeet on Windows graphics cards. A 9 minute recording takes 5.2 seconds instead of 7.4.
• Parakeet on PCs without a supported graphics card. Up to 5 times faster on processors with many cores.
• Nemotron on Windows. Recordings and Library files run on the graphics card, taking 2.8 seconds instead of a minute and a half. Live dictation feels the same.
• Nemotron on Mac. A 9 minute recording takes 6 seconds instead of 47.

Availability

Glimpse 1.3.1 is available now for Mac and Windows. If Parakeet TDT V3 is your model, Glimpse downloads the new version once in the background and keeps working while it gets ready. Macs on macOS 14 get a version tuned for that release. Dictation remains free and unlimited, with every model.