NVIDIA posts an open transcription model, 4 billion parameters active, at 5.33 percent word error

Important

Primary

On October 28, 2026, NVIDIA released Nemotron 3.5 Transcribe, an open-weight speech model. About 4 billion of its 30 billion parameters are active. On the Hugging Face English leaderboard, among open-weight models, the average word error is 5.33 percent. The post does not name the license.

At a glance

Sizes and error rates are from NVIDIA's October 28 post. The English ranking is among open-weight models.
Size	About 30 billion parameters, about 4 billion active
Languages	40 transcribed, 31 translated with English, speakers labeled in 13
English word error, open weights	5.33 percent average
Artificial Analysis speech index	2.8 percent word error, tie for first among open weights
Speaker error, 21 languages	18.26 percent, lowest of 14 systems
Granary v2 manifest	463,062 hours, 48 languages

How to read this claim

Measurement — The post calls the weights open and does not name the license.
Bill — English word error, open weights: 5.33 percent average
Door — The post calls the weights open and does not name the license.

Keep the habit · field card with prices and worked checks

What changed

NVIDIA's earlier Nemotron posts were agent and vision models. This one is an open speech checkpoint that transcribes, labels speakers, and translates.

What it introduces

An open-weight speech model with about 4 billion parameters active, covering transcription, speaker labels, and translation in one checkpoint.

Who benefits Anyone transcribing meetings or calls on their own hardware, and anyone comparing an open speech model with a paid API on a chart that left the APIs out.

Try this

If the audio is a recording rather than a live stream, start from this checkpoint and treat 5.33 percent as the open-weight English average. Check the license on the model card before you ship it. The quantized file NVIDIA timed is not the file it says you can download today.

What we do not know

The post calls the weights open and does not name the license.
The English ranking excludes API models.
The model it ties on Artificial Analysis's speech index is not named.
The 100-hour speed test is for a quantized file NVIDIA says is still forthcoming.

Evidence

Primary Model post. The size, the languages, and the error rates are from NVIDIA's October 28 post on Hugging Face.

Independent The leaderboard placements are NVIDIA's reading of public boards and of Artificial Analysis. This note does not rerun the audio.

Sources

Model post Primary · October 28, 2026

The record

What happened

NVIDIA published Nemotron 3.5 Transcribe on October 28, 2026. The checkpoint is a speech encoder plus a hybrid language model, about 30 billion parameters with about 4 billion active. NVIDIA says it is first among open-weight models on the Hugging Face Open ASR Leaderboard for English, at 5.33 percent average word error, and that API models are excluded from that ranking. A seven-language mean is 4.62 percent, and a long-form mean is 10.98 percent. On Artificial Analysis's speech index, NVIDIA says it ties for first among open-weight models at 2.8 percent word error. On that index's held-out voice-agent set, which is half the index, the error is 1.8 percent, against 2.3 percent for the next open-weight model, which NVIDIA does not name. Speaker error on David AI's 21-language track is 18.26 percent, the lowest of 14 systems, just ahead of NVIDIA's own earlier diarization model at 18.38 percent. On VoiceArena's English benchmark, NVIDIA says it is first of 14 systems at 32.4 percent concatenated word error. Granary v2, the public data manifest, covers 463,062 hours across 48 languages. NVIDIA says those hours are the dataset release, not the model's full training mix. A medical fine-tune cut held-out word error from 22.44 percent to 12.17 percent, a 45.8 percent relative drop, on NVIDIA's case study. A speed figure for 100 hours of audio uses a quantized file NVIDIA says is not out yet.

Why it matters

Speech models have been splitting the transcript, the speakers, and the translation across separate systems. This one is a single open checkpoint, and the English error rate NVIDIA leads with excludes the paid APIs. The license is the missing line on the post.

Audio question answering is trained in English. Language support is not the same for transcription, translation, and speaker labels.

The model is for recorded audio. NVIDIA says it is not a live stream.

On general English, a medical fine-tune moved LibriSpeech test-clean from 1.54 percent word error to 1.69 percent. The specialist gain cost a little general accuracy.