Gemma 4 12B it
Google · released May 23, 2026 · google/gemma-4-12B-it
Speech to textTranscribes a recording into words
- Type
- Open weightsApache License 2.0
- Languages
- 140
- Size
- 12B
Context measured in tokens
Our take
Written Sep 14, 2026Gemma 4 12B it is Google's 12-billion-parameter speech-to-text model with weights you can download under a permissive licence. It covers 140 languages and runs fast on benchmark hardware, though its English accuracy sits near the bottom of the field on every condition we measure.
Pick this when you need broad language coverage on paper and a licence that lets you modify and redistribute freely. Use it for local deployment where Apache 2.0 matters more than measured accuracy. Skip it if you need reliable transcription of meetings, podcasts, accented speech, or clean read-aloud audio, or if you want a hosted provider to run it for you.
The case for it
- 140 languages listed, far more than most transcription models cover.
- Apache 2.0 licence allows commercial use, modification and redistribution without restriction.
- Processes an hour of audio in about 34 seconds on the benchmark harness.
The case against it
- Near-bottom accuracy on every measured English condition: roughly one word in five wrong overall, and more than three-quarters wrong in recorded meetings.
- No commercial hosting options currently available.
- No accuracy data for any language other than English.
How good is it?
An open speech-to-text model from Google that turns recordings into written text, though it trails most others on accuracy.
- turning spoken English into written textOpen ASR WER · 76th of 76
- transcribing recordings of meetings in a roomRecorded meetings · 92nd of 92
- transcribing speakers with a range of accentsAccented speech · 71st of 76
- transcribing podcasts and video audioPodcasts and video · 92nd of 92
- transcribing clear recordings of people reading aloudClean read speech · 92nd of 92
TranscriptionTurning speech into text1 of 5Open ASR WER · 76th of 76
79.1%
Misses roughly one word in 5, averaged over nine English test sets.
106×69th of 74
an hour of audio in 34 seconds, on the board's own hardware. Your machine will differ.
140
Stated by the leaderboard; we do not hold the list itself.
Percentage of words wrong on each set, lower better. Bars are scaled to this model's own worst case; the placing beneath each rate is against every model measured on that set.
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.
Each of these is the same transcription job on a different kind of recording, so together they say where it holds up and where it slips — not how closely it follows an instruction.
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.
Can you run it yourself?
Comfortable fit
GeForce RTX 4090 · 24 GB
Room to spare. 13 GB spare means a 10% error in the size would not change the answer.
GeForce RTX 5090 · 32 GB
Room to spare. 21 GB spare means a 10% error in the size would not change the answer.
Apple M1 (8-core GPU) · 16 GB
Room to spare. 2.2 GB spare means a 10% error in the size would not change the answer.
These cards answer whether Gemma 4 12B it loads, not how fast it transcribes. Throughput figures for a transcription model come from its text decoder, so treat this as a fit answer rather than a speed one.
Memory use by level
Against a 24 GB card.
What is quantisation? →This model on every device we track71 devicesThe Q4 build most people download, on each device: what the weights come to, how much context the memory leaves, and whether it runs. Smallest device that runs it first. This is a fit answer: whether it loads, not how fast it transcribes.
When we formed this view
Recent changes
What moved
first indexed by our pipelineEach date is the day we first saw the change, or the day the maker announced it.
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.
Licence and identifiers
What the licence allowsApache License 2.0, 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
Apache License 2.0
Fully permissive: commercial use, redistribution, and derivatives allowed. Requires attribution and a copy of the license. Includes an express patent grant.
Identifiers
- Hugging Face
- google/gemma-4-12B-it
- Architecture
- Dense
- Takes in, gives back
- Audio in, text out
- Catalogue slug
- google-gemma-4-12b-it