400 model_decommissioned: distil-whisper-large-v3-en (Groq)
If your dictation app, meeting recorder or voice bot stopped transcribing and the Groq response says 400 The model `distil-whisper-large-v3-en` has been decommissioned and is no longer supported, your key and your audio are fine. Groq retired its English-only Distil-Whisper model on 23 August 2025 and named whisper-large-v3-turbo as the replacement. A lot of apps shipped the id as a hard-coded preset or as the default in a settings file, so users keep hitting it more than a year later. In many apps it fails silently: the recording works, then nothing is typed.
Last verified November 21, 2026 against Groq deprecations, Groq speech-to-text docs, Error body reported in dimmy#63 · 5 min read
| Provider | Groq |
| Endpoint | POST /openai/v1/audio/transcriptions |
| HTTP status | 400 Bad Request |
| Error code | model_decommissioned |
| Message | The model `distil-whisper-large-v3-en` has been decommissioned and is no longer supported. Please refer to https://console.groq.com/docs/deprecations for a recommendation on which model to use instead. |
| Announced / shut down | 23 July 2025 / 23 August 2025 |
| Groq's replacement | whisper-large-v3-turbo |
| Can you wait it out? | No. The model was retired; every request fails until you change the model or the provider |
- Fastest: change the model string to
whisper-large-v3-turboand stay on Groq. Same request shape, multilingual, slightly slower than Distil was. - Least code change: keep
distil-whisper-large-v3-enand change only the base URL; Inference APIs accepts the id and transcribes with Whisper Large v3 Turbo at $0.001/min. - For shipped apps: migrate saved settings that still hold the old id at load time, and show the provider's error text instead of failing silently.
What happened
Groq's deprecations page lists the notice: announced 23 July 2025, shut down 23 August 2025, all self-serve tiers, replacement whisper-large-v3-turbo, "a more performant model for speech recognition and transcription tasks, and supports more languages". Distil-Whisper was a distilled, English-only version of Whisper Large v3, popular because it was the fastest and cheapest option on Groq's list.
Groq answers retired ids with HTTP 400 and the code model_decommissioned, not with a 404. Clients that only handle 401, 413 and 429 often treat it as a generic failure, which is why the symptom in dictation apps is usually "nothing happens" rather than an error dialog. Older Groq retirements behave the same way; the deprecations tracker lists them all.
First, find where the id is pinned
In dictation and transcription apps the id usually lives in three places: a preset list in the UI, the default value in the settings model, and the settings file already saved on every user's machine. Fixing the first two does not fix the third.
grep -rn "distil-whisper-large-v3-en" . --include='*.py' --include='*.ts' --include='*.js' --include='*.swift' --include='*.kt' --include='*.rs' --include='*.cs' --include='*.go' --include='*.json' --include='*.yaml' --include='*.toml' --include='*.env*'
# and the copies grep cannot see: users' saved settings files, environment variables on your serverOption 1: switch to Groq's replacement
One string changes and nothing else in the request does. What is different about whisper-large-v3-turbo:
- It is multilingual. Distil was English-only, so it never guessed the language. Turbo detects it; if your users dictate only English, send
language=enso a short or noisy clip is not transcribed as another language. - Accuracy is similar or better on English, and much better on accents and mixed-language speech.
- It is a little slower per request than Distil was. For push-to-talk dictation the difference is a fraction of a second.
- The free tier still meters audio seconds per hour and per day. Long recordings and heavy dictation users hit
429with "audio seconds per hour"; see the Groq 429 page.
resp = client.audio.transcriptions.create(
model="whisper-large-v3-turbo", # was "distil-whisper-large-v3-en"
file=open("clip.m4a", "rb"),
language="en", # Distil was English-only; keep that behaviour
)Option 2: keep the old id, change the base URL
Inference APIs accepts distil-whisper-large-v3-en (and whisper-large-v3-turbo, whisper-1, whisper-large-v3) on an OpenAI-compatible /v1/audio/transcriptions endpoint. The retired id is served by Whisper Large v3 Turbo running on our own GPU, so saved settings that still hold the old id start working again without an app update. An 8-second dictation clip came back in 0.7 s end to end when we measured it on 21 November 2026.
- Price: $0.001 per audio minute, billed per request from a prepaid balance. OpenAI charges $0.006. Groq's paid tier is cheaper per minute ($0.04 per hour), but its free tier caps audio per hour and per day; we do not.
- Formats: json, text, srt, vtt, verbose_json with word and segment timestamps; m4a, mp3, wav, webm, ogg, flac.
- Same key also serves the LLM cleanup pass most dictation apps run after transcription (DeepSeek, Llama, Qwen, GPT-OSS).
from openai import OpenAI
client = OpenAI(
base_url="https://api.inferenceapis.com/v1", # was https://api.groq.com/openai/v1
api_key=os.environ["INFERENCE_API_KEY"],
)
resp = client.audio.transcriptions.create(
model="distil-whisper-large-v3-en", # the retired Groq id, accepted as an alias
file=open("clip.m4a", "rb"),
)
print(resp.text)curl https://api.inferenceapis.com/v1/audio/transcriptions \
-H "Authorization: Bearer $INFERENCE_API_KEY" \
-F model=distil-whisper-large-v3-en \
-F file=@clip.m4aApps with an "OpenAI compatible" transcription setting need only the base URL and key; guides for VoiceInk and TypeWhisper show where they go. Details and benchmarks: speech-to-text API, Whisper Large v3 Turbo.
Option 3: for app developers, migrate saved settings
If you ship an app, your users' saved settings still contain the old id after you update the preset list. Rewrite it when settings load, and surface the provider's error message so the next retirement is visible on day one instead of looking like a broken microphone.
const RETIRED_STT: Record<string, string> = {
"distil-whisper-large-v3-en": "whisper-large-v3-turbo", // Groq, 2025-08-23
};
function loadSettings(raw: Settings): Settings {
const next = RETIRED_STT[raw.sttModel];
return next ? { ...raw, sttModel: next } : raw;
}
// and when a request fails, show the body, not just "transcription failed"
if (!res.ok) throw new Error(`${res.status}: ${(await res.json()).error?.message}`);Letting users point the transcription provider at any OpenAI-compatible base URL also makes the next retirement a settings change instead of a release.
Frequently asked questions
Does anyone still host distil-whisper-large-v3-en?
Not on Groq. The open weights (distil-whisper/distil-large-v3) can be self-hosted, and Inference APIs accepts the Groq id and transcribes it with Whisper Large v3 Turbo, so existing configurations keep working with only a base-URL change.
Why does my app fail silently instead of showing an error?
Groq returns HTTP 400 with code model_decommissioned. Many clients only special-case 401, 413 and 429 and swallow everything else. Log or display the error.message field of the response body.
Is whisper-large-v3-turbo English-only like Distil was?
No, it is multilingual and detects the language. Pass language=en if your users only dictate English, otherwise very short clips are occasionally transcribed as another language.
How much does transcription cost after the switch?
Groq bills whisper-large-v3-turbo at $0.04 per hour on its paid tier and meters audio seconds on the free tier. Inference APIs bills $0.001 per minute ($0.06 per hour) with no audio cap; OpenAI whisper-1 is $0.006 per minute.
Where Inference APIs fits
Option 2 above keeps the old id working with a base-URL change: Inference APIs accepts distil-whisper-large-v3-en on /v1/audio/transcriptions and serves it with Whisper Large v3 Turbo on our own GPU at $0.001 per audio minute, billed per request with no hourly or daily audio cap.
Something changed or wrong? Tell us and we will re-verify the entry.
