whatmodel.sh / stt / Kyutai STT 2.6B (en)

Kyutai STT 2.6B (en)

+ contribute
Update model informationAdd a place to run itAdd a test resultPropose a benchmarkCorrect a benchmark scoreEdit data on GitHub

Kyutai STT 2.6B (en) is an open-weights speech-to-text model from Kyutai, released June 2025. Its published size is 2.6B parameters. Licence: cc-by-4.0.

56.9±3.5#57 of 162 speech-to-text modelsfrom 3 benchmark results across 2 sources

Score within the Open ASR English short-form WER (March 2026) comparison set only.

$ how it compares

The speech-to-text models ranked either side of Kyutai STT 2.6B (en), out of 176 with a score.

#modelscorewhereRAMreleased
64Parakeet Redux Moondream57±4local~0.2GB2026-09
65Parakeet TDT 0.6B v3 NVIDIA57±1api $0.0015/audio min +local~2GB2025-08
66Kyutai STT 2.6B (en) Kyutai57±4local~6.7GB2025-06
67Whisper large-v3 OpenAI57±1api $0.0018/audio min +local~4.2GB2023-11
68GLM-ASR-Nano-2512 Z.ai57±4local~5.9GB2025-12

see all speech-to-text models · find the ones that fit your machine

$ what it's good at

long-form transcription█·····█···██ #4 of 4 * · 26% of its score · 1 test

Trails Voxtral Mini 4B Realtime (2602), GPT-4o Mini Transcribe here, though they rank lower overall.

testresultvs the others in its rundifficultycountswhy
Voxtral report · English long-form 4-set WER7.7% ✎▄··█·······▄ beats 2/3avg 11%▇▇▇▇▇▇▇▇▇▇ 26%typical
other tests2 not in a task · 69% of its score
testresultvs the others in its rundifficultycountswhy
Open ASR English short-form WER (March 2026)6.4% ✎▁▁▁▃··▁▁▂██▆ beats 72/86avg 8.2%▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇ 42%blunt
Voxtral report · English short-form 10-set WER8.1% ✎█··█·█·····█ beats 1/3avg 12%▇▇▇▇▇▇▇▇▇▇▇ 27%typical

$ where to run it

Other ways to run it

version

  1. Run it as published · needs ~6.74 GB, NVIDIA only · from github.com/kyutai-labs
    uvx --with moshi python -m moshi.run_inference --hf-repo kyutai/stt-2.6b-en audio.mp3

Official weights and documentation: huggingface.co/kyutai ↗

$ how the score was worked out · Voxtral Realtime 53%, Open ASR Leaderboard 42%

Each result implies a score on its own, after adjusting for how hard its test is and who ran it. The score is their weighted average, with a small pull toward the middle for thinly tested models. Each test's weight is its counts column above; here it's added up by who ran the tests. How it works

sourceresultsshare of the scorelicence
Voxtral Realtime2▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇ 53%CC BY 4.0
Open ASR Leaderboard1▇▇▇▇▇▇▇▇▇▇▇▇▇ 42%Apache-2.0
prior · pull toward the middle▇ 5%
weighted average= 56.9

$ improve this page

Update model informationAdd a place to run itAdd a test resultPropose a benchmarkCorrect a benchmark scoreEdit data on GitHub

Use a short GitHub form or edit the data through a pull request. Every number needs a source; here's which ones we can use.

data updated · sources and licences · json