Tasks and inputs
Nemotron 3.5 Transcribe 30B A4B is listed for Automatic speech recognition, Translation, Speaker Diarization. It accepts audio and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.
SyncDev model directory · Profile updated Nov 17, 2026
Nemotron 3.5 Transcribe 30B A4B is NVIDIA's open offline ASR model: 40 languages, hour-long files, word timestamps, diarisation and translation.
The specification score summarizes recorded features and limits. It is not a benchmark result.
Pricing conditions: Weights free under OpenMDW-1.1. OpenRouter lists an audio-based price; confirm the unit and rate on your account.
Recorded modality support shows what this model can accept and produce. An undocumented modality is shown as unknown rather than assumed to be unsupported.
Accepts
Produces
An hour of audio in one request, with words, timestamps and speaker turns coming back together. Nemotron 3.5 Transcribe 30B A4B is NVIDIA's open-weight offline speech model, and its pitch is that one checkpoint replaces a recogniser, an aligner and a diarisation pipeline. NVIDIA dates it 28 October 2026.
It is not a streaming model. Audio is encoded internally in 30-second chunks, and the whole file goes in at once.
The language model is a hybrid Mamba-2/Transformer mixture-of-experts built on Nemotron 3.5 Lightning 30B A3B. Alongside it runs a parallel encoder: a Transformer ASR encoder plus a speaker encoder. Totals come to about 34B parameters with about 4.3B active, including an audio encoder of roughly 800M.
| Item | Detail |
|---|---|
| Weights | nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B (BF16; FP8 and NVFP4 variants) |
| Languages | 40 for transcription; translation to and from English for a subset |
| Max recording | About one hour |
| Audio | 16 kHz mono; WAV, FLAC, MP3, OGG |
| Context | 65,536 tokens |
| Extras | Word timestamps, speaker diarisation |
| Licence | OpenMDW-1.1 |
NVIDIA reports a 4.71% mean word error rate on the nine-dataset English Open ASR Leaderboard and 4.62% across its seven-language multilingual set, ahead of Qwen3-ASR-1.7B on both. On long-form English it reports 10.98 against 13.37 for Parakeet-TDT-0.6B-v3. These are publisher numbers.
This is fiddly, so plan for it. In vLLM, start the server with the ctc_timestamp_middleware and --trust-request-mm-kwargs, then send capture_ctc_timestamps: True in the multimodal processor kwargs, with speaker tags such as <spk:0> in the prompt. The response carries a ctc_timestamps object holding words, timings, diarisation segments and the speaker map. Always check its error field.
Constraints: one recording per request, stream=false, n=1. And timestamps do not work with the DFlash2 speculative-decoding option.
For plain transcription, NVIDIA recommends temperature 0 with reasoning off.
NVIDIA's NIM, with NVFP4 and DFlash2 on one RTX PRO 6000 Blackwell, reportedly processes up to 2,238 seconds of audio per second. Those serving optimisations are not open source, and the open vLLM path is expected to be about half as fast. Budget for the open number if you self-host.
The reference deployment is one H100 80GB. Lovelace cards such as the L40S run FP8 only.
The Hugging Face repo was created on 7 October, three weeks before the stated release date, and OpenRouter added it on 29 October as nvidia/nemotron-3.5-transcribe with an audio-based price whose unit we could not confirm. NVIDIA mentions a Transcribe NIM, but the model did not appear in the build.nvidia.com catalogue when checked. Verify hosted availability before you plan around it.
Benchmark it head to head with Whisper Large v3 Turbo on twenty of your longest, messiest files. Score WER, then score diarisation separately with cpWER or a manual check of speaker turns. Timestamp drift at the end of an hour-long file is the thing to look at closely.
More in automatic speech recognition and on NVIDIA's provider page.
Compare the model on the work you need to complete.
Nemotron 3.5 Transcribe 30B A4B is listed for Automatic speech recognition, Translation, Speaker Diarization. It accepts audio and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.
Documented limits: 65.54K context. Measure a completed task using the prompts, documents and tools your application will use. Include reasoning and retries in the result.
The weights are downloadable under OpenMDW-1.1. Check runtime support and size the complete model and working context for your expected concurrency.
These records share a directory category. Check the source card, output type and deployment requirements before treating a model as a replacement.
salesforce
SOV-Transcribe-1 is Salesforce's open-weight speech-augmented model for non-streaming transcription, with keyword biasing and style instructions; Apache-2.0.
ElevenLabs
Scribe v3 Live is ElevenLabs' streaming transcription model with turn prediction, mid-sentence language switching and a 10,000-term keyterm library.
IBM
Granite Speech 5.0 2B is IBM's open speech-language model for multilingual ASR and translation with keyword biasing, runnable from Ollama to MLX.
Mistral AI
Voxtral Mini Transcribe 3 is Mistral's 4B batch transcription model with speaker labels, segment timestamps and custom vocabulary, replacing Mini 2.
Mistral AI
Voxtral Mini Transcribe Realtime 3 is Mistral's streaming speech-to-text model for voice agents: under 500 ms, tunable delay, built-in turn detection.
Mistral AI
Voxtral Small Transcribe 3 is Mistral's top-accuracy batch speech-to-text model for recorded audio, with context biasing and 50+ languages.