Flit 1.6: four new voice models, and more than 1,600 languages
In short
Flit 1.6 adds four voice models, and every one runs on your Mac. Parakeet Ultra makes fewer mistakes than Parakeet V3 in English and 24 other European languages, at the same speed, and is now the model Flit recommends for them. Voxtral Realtime writes while you speak in 13 languages, Hindi, Arabic, Japanese and Korean among them. Omnilingual ASR covers more than 1,600 languages, so Tamil, Bengali, Urdu and Swahili have a model too. Parakeet Unified is a new English model. The model you use now stays as it is, and 1.6 is a free update for every license.
Which model to use
You only need one voice model. Choose it by the language you dictate in:
- English or another European language: Parakeet Ultra 0.6B. Your words appear while you speak, and it is the fastest model you can download. For German, French, Spanish, Italian or Dutch on a Mac with 16 GB of memory or more, Cohere Transcribe is more accurate, and its text appears when you stop speaking.
- Hindi, Arabic, Japanese, Korean or Chinese, with words while you speak: Voxtral Realtime 4B.
- A language none of these cover, such as Tamil, Bengali, Urdu, Marathi or Swahili: Meta Omnilingual ASR 3B.
- English only, and you want to compare: Parakeet Unified 0.6B.
Flit suggests one when you set it up, from the languages you say you speak, and Settings → Models shows the same suggestion at the top.
Download a model and switch to it
The new models are a download away, and the one you use now keeps working while they download.
- Right-click the Flit icon in the menu bar and choose Settings…, then Models in the sidebar.
- Find the model in the Library and click Download.
- When it says Downloaded, click the row. It is marked Active.
- Dictate as usual. The first dictation after a switch takes a few seconds longer while the model loads.
The bin on a row deletes a model you no longer use and gives the space back.
Parakeet Ultra: fewer mistakes, the same speed
Parakeet Ultra is Parakeet V3 trained further by moondream. It covers the same 25 languages and runs at the same speed on the Neural Engine, with fewer mistakes: across 24 languages of the FLEURS test set, its average word error rate is 11.7%, against 14.8% for V3.
In our own test, a Danish sentence read aloud came back nearly word for word, where V3 heard "torsdag eftermiddag" as "torsdag efter med dag". Your words still appear while you speak, and the rest lands about a tenth of a second after you let go of the key.
If you use Parakeet V3 today, nothing changes until you switch. V3 stays in the list.
A 595 MB download, about 100 MB more than V3.
Voxtral Realtime: live words in 13 languages
Voxtral Realtime, from Mistral, writes while you speak, punctuation included, in English, French, Spanish, German, Russian, Chinese, Japanese, Italian, Portuguese, Dutch, Arabic, Hindi and Korean. It works out which of them you are speaking, so there is no language to set.
Words appear a second or two behind your voice, and the rest lands within half a second of letting go of the key. In our tests, Hindi and Arabic sentences came back nearly word for word.
It is the largest of the four: a 2.8 GB download that holds a few gigabytes of memory while it runs, so it suits a Mac with 16 GB or more.
Omnilingual: more than 1,600 languages
Meta's Omnilingual ASR hears more than 1,600 languages, among them Tamil, Telugu, Bengali, Marathi, Urdu, Punjabi, Swahili, Yoruba and Amharic. If no other model covers a language you speak, Flit recommends it when you set up.
It writes plain words, with no punctuation or capitals. Flit's formatting adds them at Polished, the level in Settings → Formatting. At Verbatim you get the words exactly as the model wrote them.
Accuracy depends on the language. In our tests it was close to word for word in Hindi, made a few slips in Arabic, and more in Tamil and Bengali. Where another model covers your language, that model is the better choice.
A 1.8 GB download. Text appears when you stop speaking.
Parakeet Unified: a new English model
Parakeet Unified, from NVIDIA, is trained for live words and whole recordings at once. It is English only, shows your words while you speak, and is a 570 MB download. If you only dictate in English, try it beside Parakeet Ultra and keep the one you like.
Use them for recordings too
Every new model works in the Transcribe tab. Choose it under Model below the card, and Flit uses it for files while dictation keeps its own model. That is how a recording in Tamil, Hindi or Korean becomes a transcript on your Mac.
Questions people ask
Is Flit 1.6 a free update?
Yes. Every license includes twelve months of updates, and 1.6 is one of them. The trial has every model too.
How do I get it?
Right-click the Flit icon in the menu bar and choose Check for Updates…. Your model, settings, dictionary and history carry over, and a window after the restart shows what is new.
Does my voice model change?
No. Flit keeps the model you use until you choose another. The new ones are a download away in Settings → Models.
Which model is best for my language?
Open Settings → Models and add your languages under Languages you speak. The card at the top names the model Flit recommends for them, and why.
Do I need all of them?
No. One voice model is enough. Each is a download of 570 MB to 2.8 GB, and you can delete the ones you do not use.
Is anything uploaded?
No. Every model runs on your Mac, and Flit dictates with Wi-Fi off. The only network request is the download you start yourself.
Why does Omnilingual write without punctuation?
The model writes the words only. Flit's formatting adds punctuation and capitals at the Polished level.
Does Voxtral work on a Mac with 8 GB of memory?
It is built for 16 GB or more. Beside the formatting model it leaves little memory for your other apps, so on 8 GB, Parakeet Ultra or Qwen3-ASR 0.6B is the better choice.
Everything in 1.6.0
Flit 1.6 is a free update. Right-click the Flit icon in the menu bar and choose Check for Updates. New to Flit? It is free for 7 days with every feature on, then $19 once.
- New Parakeet Ultra. Parakeet V3 trained further: the same 25 languages and the same speed, with fewer mistakes. Your words appear while you speak, and it is now the model Flit recommends for English and 24 other European languages.
- New Voxtral Realtime. Mistral's model writes while you speak, punctuation included, in 13 languages: English, French, Spanish, German, Russian, Chinese, Japanese, Italian, Portuguese, Dutch, Arabic, Hindi and Korean.
- New More than 1,600 languages. Meta's Omnilingual ASR hears languages no other model in Flit does, such as Tamil, Bengali, Urdu, Marathi and Swahili. It writes plain words, and Flit's formatting adds the punctuation.
- New Parakeet Unified. NVIDIA's newest English model, trained for live words and whole recordings at once.
- Better Setup recommends Omnilingual for a language no other model covers, where it used to fall back to Apple's built-in engine. Your current model stays as it is until you choose another.