Models / Kyutai/ STT 2.6b en

STT 2.6b en

Kyutai · released Jun 6, 2025 · kyutai/stt-2.6b-en

Speech to textTranscribes a recording into words

Input: audio. Output: text.InputOutput
Type
Open weightsCreative Commons Attribution 4.0
Languages
1
Size
2.6B

Context measured in tokens

Our take

Written Aug 3, 2026

STT 2.6b en is a compact downloadable English speech-to-text model from Kyutai that processes audio extremely fast on benchmark hardware. Its accuracy is excellent for clean read-aloud recordings but drops sharply on accented speech or meetings, and it must be self-hosted as no providers currently offer it.

Who should pick it

Choose this for local transcription of clean, single-speaker English audio where speed matters — it processes an hour of audio in under half a minute on benchmark hardware. Use it for research or integration projects that need open weights under a permissive attribution licence. Skip it if you need hosted inference, work with accented speakers or meeting recordings, or require language coverage beyond English.

The case for it

  • Processes audio at 133 times real time on benchmark hardware — one hour of audio in 27 seconds.
  • Strong accuracy on clean read-aloud speech, with about one word in seventy wrong.
  • Creative Commons Attribution 4.0 licence allows commercial use with attribution.

The case against it

  • Accuracy degrades sharply on challenging audio: more than seven times worse on accented speech and recorded meetings.
  • No hosted inference options available; users must self-host.
  • English only, with no verified data on performance in other languages.
00

How good is it?

TranscriptionTurning speech into text3 of 5Open ASR WER · 36th of 74

Words it gets right

94.3%

Misses roughly one word in 17, averaged over nine English test sets.

How fast it listens

133×56th of 62

an hour of audio in 27 seconds, on the board's own hardware. Your machine will differ.

Languages

1

Listed on the model card. The accuracy above is English only.

Where it struggles
Read aloudaudiobooks, clean recording1.4%
Podcasts and videoeveryday internet audio8.2%
Accented speechspeakers from many countries10.6%
Meetingsa room, several people, far microphone10.6%

Percentage of words wrong on each set, lower better. Bars are scaled to this model's own worst case, not to the board.

The figures above come from the Open ASR Leaderboard, an independent public test that runs every model on the same recordings. It is the only measurement of transcription quality we know of, so there are no other scores to show.

Also scored, on boards we give no mark for
Financial calls 13th of 74European-accented speech 29th of 74Accented speech 30th of 74Recorded meetings 36th of 74Podcasts and video 37th of 74Clean read speech 37th of 74Harder read speech 51st of 74

These tests check whether a model follows instructions — a precondition for all the work above, but not a measure of how well that work is done, which is why they get no rating.

Every published score for this model9 scoresEvery figure we hold, from 9 boards, with who ran it and a link to the source — including the boards no rating above is built on.
133.2independentsource ↗
5.7independentsource ↗
10.6independentsource ↗
1.9independentsource ↗
10.6independentsource ↗
8.2independentsource ↗
1.4independentsource ↗
01

Can you run it yourself?

Fits in memory
weights load entirely on the card
Spills to system RAM
some weights offload; much slower
Too large
will not load even with offload
est
size is calculated; the verdict could change by 10%

Comfortable fit

A card many people ownFits in memory

GeForce RTX 4090 · 24 GB

Weights at Q4_K_M1.6 / 24 GBest
Spare memory19.9 GB spare
Usable context262Kwhat the spare memory holds; no published limit on record
Decode speed512 tok/sest

Room to spare. 19.9 GB spare means a 10% error in the size would not change the answer.

One step upFits in memory

GeForce RTX 5090 · 32 GB

Weights at Q4_K_M1.6 / 32 GBest
Spare memory27.9 GB spare
Usable context262Kwhat the spare memory holds; no published limit on record
Decode speed910 tok/sest

Room to spare. 27.9 GB spare means a 10% error in the size would not change the answer.

On a MacFits in memory

Apple M2 (8-core GPU, 8GB unified) · 8 GB

Weights at Q4_K_M1.6 / 8 GBest
Spare memory3.1 GB spare
Usable context66Kwhat the spare memory holds; no published limit on record
Decode speed44 tok/sest

Room to spare. 3.1 GB spare means a 10% error in the size would not change the answer.

Your hardware
Checking your profile…

Memory use by level

Against a 24 GB card.

Q4_K_M
recommended
1.6 GBest
Fits in memory
Q5_K_M
1.9 GBest
Fits in memory
Q8_0
2.9 GBest
Fits in memory

All 3 sizes here are calculated, not measured. We hold no measured file for this model, so each size comes from the parameter count and every verdict above inherits that uncertainty.

Check against your own machine → · All 71 devices, with every size →

02

When we formed this view

Dates behind this page

Aug 2, 2026BenchmarkScored 10.6 on Accented speechleaderboard
Aug 2, 2026BenchmarkScored 1.9 on Financial callsleaderboard
Aug 2, 2026BenchmarkScored 10.6 on Recorded meetingsleaderboard
Aug 2, 2026BenchmarkScored 3.5 on European-accented speechleaderboard
Aug 2, 2026BenchmarkScored 8.2 on Podcasts and videoleaderboard
Aug 2, 2026BenchmarkScored 1.4 on Clean read speechleaderboard
Aug 2, 2026BenchmarkScored 4 on Harder read speechleaderboard
Aug 1, 2026ListedListed on LLMapfirst indexed by our pipeline
Aug 1, 2026BenchmarkScored 133.2 on Open ASR RTFxleaderboard
Aug 1, 2026BenchmarkScored 5.7 on Open ASR WERleaderboard

Prices last checked 6h ago

What we do not know about this model yet

  • We hold no measured file for it, so all 3 sizes on this page are calculated from the parameter count.
  • Nothing we hold says whether an endpoint streams, so we do not show it either way.
03

Licence and identifiers

What the licence allowsCreative Commons Attribution 4.0, what it allows commercially, and the identifiers you need to pull this model — its Hugging Face repo, our slug and a machine-readable card.

Licence

Creative Commons Attribution 4.0

permissiveCommercial use allowed

Permissive content license: any use with attribution. Common for datasets and some model weights.

Identifiers

Hugging Face
kyutai/stt-2.6b-en
Architecture
Dense
Modality record
audio->text
Catalogue slug
kyutai-stt-2-6b-en

Machine-readable model card (omc.json) →

Something wrong on this page? Tell us