NVIDIA Nemotron 3.5 Transcribe is an open-weight speech and audio model for offline processing. It combines transcription with instruction-conditioned tasks such as contextual transcription, speaker attribution, translation and audio questions. It is a different product surface from a text-only Nemotron coding model.
The durable decision is whether its language coverage, task controls and deployment footprint fit your audio pipeline. This is a documentation-based guide. We have not measured its error rate, throughput or speaker attribution on independent recordings. The official model card is the starting point for current runtime and task-specific requirements.
What is available
| Property | Published specification |
|---|---|
| Main checkpoint | nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B |
| Processing | Offline speech and audio to text |
| Languages | 40, with task-specific support differences |
| Recording length | Up to one hour |
| Context | 65,536 tokens |
| Active parameters | Approximately 4.3B, including audio encoder |
| Variants | BF16, FP8, NVFP4 and GGUF routes |
| License | OpenMDW-1.1 |
The name includes 30B-A4B, but NVIDIA’s descriptive active-parameter count includes the audio encoder. Preserve the published counting convention rather than trying to make the label and the complete system identical.
The research page links the release artifacts. The runtime documentation has additional deployment details. A nightly documentation route may evolve; pin the runtime and checkpoint used in your evaluation.
Transcription, diarization and translation are separate jobs
Ordinary transcription asks what was said. Contextual transcription can include supporting terms or instructions. Speaker-attributed transcription also asks which speaker said it. Translation changes the output language. Audio question answering asks for an interpretation of the recording rather than necessarily returning a verbatim transcript.
Those tasks need different acceptance tests. A readable summary can omit a name or amount that a transcription workload must preserve. Correct words do not prove correct speaker labels. Translation quality should be checked separately from the source-language transcript.
Consult the model card’s language support matrix for the exact task. The headline 40-language count is not a guarantee of equal accuracy or identical features in every language. Include the accents, noise, microphones and vocabulary your application will encounter. Our speech-model comparison provides context for comparing APIs and local pipelines without treating every speech task as the same benchmark.
GPU requirements depend on precision and implementation
NVIDIA documents a single H100 80 GB deployment for the main model. Its NVFP4 profile targets Blackwell hardware, including an RTX PRO 6000 Blackwell 96 GB example. These are supported profiles, not independently established minimum requirements for every runtime.
Do not assume that a quantized checkpoint is compatible with any GPU that has enough memory. Precision formats and optimized kernels have architecture requirements. Check the hardware table before choosing BF16, FP8 or NVFP4. A GGUF package is another implementation path with its own supported operations.
A useful evaluation begins with short audio, then increases recording length and concurrency separately. Record peak memory, preprocessing time, queue time and output latency. Include the transcript format and task in each result so that later comparisons remain meaningful. The GPU memory guide explains why weights alone do not determine a server budget.
One hour does not mean streaming
This release is described as offline processing. Support for a one-hour recording does not establish a low-latency streaming interface. A near-live application still needs a documented streaming path or a carefully designed chunking strategy.
Chunking can change speaker attribution, timestamps and words spanning a boundary. If you divide a recording, define overlap, reconciliation and segment identifiers. Preserve the original audio timing so a reviewer can trace a disputed transcript back to the source. Do not claim full-recording coherence from a set of disconnected chunks.
The 65,536-token context is also a capacity limit, not a guarantee that every hour of dense audio plus instructions fits every serving setup. Test the actual media duration, channel layout and output task together.
Runtime integration and task settings
Use NVIDIA’s linked NeMo, Transformers or optimized serving instructions for the chosen artifact. A generic OpenAI-compatible client is not proof that an arbitrary speech endpoint exposes the same controls. Verify the accepted audio packaging, prompt format, output schema and task parameters.
NVIDIA recommends deterministic settings for transcription, translation and language identification, with reasoning disabled. Its audio-understanding recommendation uses different sampling and reasoning settings. Keep those profiles separate rather than applying a text-chat default everywhere.
Before production, test malformed audio, unsupported languages, empty recordings, interrupted requests and output parsing. Preserve failures rather than silently replacing them with an empty transcript. The regression testing guide and AI testing foundation cover repeatable evaluation records.
Pricing and license
Released weights do not establish a public per-minute API price. No independent current hosted tariff is asserted here. Local cost includes the GPU, idle capacity, storage, preprocessing, operating support and retries. Compare cost per accepted recording, not only the time spent in model inference.
NVIDIA releases this checkpoint under OpenMDW-1.1. Review the full model terms and the licenses of the serving packages. A self-hosted system also needs its own audio retention, access controls and deletion policy. Local inference does not automatically remove logs, backups or downstream cloud services. See the data-retention guide for that distinction.
My take
This model is a serious candidate when you need instruction-controlled offline transcription and can operate the documented GPU stack. Evaluate speaker attribution and specialized vocabulary separately from ordinary word accuracy. For short or low-volume jobs, compare a hosted service before committing to a dedicated GPU. NVIDIA’s benchmark and speed claims are useful hypotheses to test, not measurements of your recordings.
FAQ
Is Nemotron 3.5 Transcribe a streaming model?
The documented release is offline. Do not infer streaming latency from a supported recording duration or vendor throughput result.
Does it support Dutch?
Dutch is listed among the languages, but consult the task-specific support matrix and evaluate your own recordings. Language inclusion does not establish accuracy for a domain or accent.
Will NVFP4 run on any NVIDIA GPU?
No. Check the precision and hardware compatibility table. The documented NVFP4 profile targets Blackwell hardware.
Are the published speed results our tests?
No. They are NVIDIA-reported results under its stated hardware and serving configuration.