Latest in AI
NVIDIA Nemotron 3.5 Transcribe: Languages, API Access and Deployment Limits
NVIDIA's new speech model takes instructions as well as audio, but its launch claims need a careful distinction between recorded speech and live voice.
NVIDIA released Nemotron 3.5 Transcribe on October 28, an open-weight model for turning recorded speech into transcripts, translations and answers. Developers can tell it how to handle terminology and formatting, which makes it relevant to applications where a technically correct transcript still needs substantial cleanup.
The useful distinction is recorded audio, not live streaming. A fast batch transcription result does not mean this checkpoint is ready to handle a real-time voice conversation.
What Nemotron 3.5 Transcribe supports
NVIDIA's launch explanation describes transcription in 40 languages, translation between English and 31 languages, and speaker attribution in 13. Recordings can contain up to eight speakers. Those are different coverage numbers, so a language being supported for transcription does not mean every other task works in it.
For a meeting product, this lets the same recording support several requests: produce a readable transcript, identify speaker turns, or answer a question about the discussion. The instructions can also include product names or technical vocabulary.
That changes where developers can customize the pipeline. Instead of relying entirely on a cleanup model after transcription, they can provide context before recognition. The tradeoff is that a suggested term can influence the output. Vocabulary hints should be evaluated against what was actually said, rather than treated as a guarantee of accuracy.
Model size and hardware requirements
The model card gives the following specifications:
| Detail | Published specification |
|---|---|
| Full model, including audio encoder | Approximately 34 billion parameters |
| Active parameters, including audio encoder | Approximately 4.3 billion |
| Context | 65,536 tokens |
| Recording length | Up to one hour |
| Reference deployment | One H100 with 80 GB memory |
| License | OpenMDW-1.1 |
The 30B-A4B name should not be read as the full memory requirement. Active parameter count describes the portion used in computation, not the amount of weight storage a deployment needs.
There is also a launch-documentation mismatch worth checking before planning a smaller local deployment. The announcement says quantized versions are coming soon, while the model card lists FP8 and NVFP4 variants as available. Confirm the particular artifact and serving support you need rather than assuming every advertised configuration is ready.
API access and pricing
You can download weights or use a hosting provider. Baseten's deployment page documents an OpenAI-compatible chat-completions request with a recording and text instruction. It targets recorded-audio workflows such as post-call analysis and meeting tools.
That compatibility is useful for teams that already have a client abstraction around chat-completions APIs. It does not make transcription interchangeable with an ordinary text request: audio encoding, output length, speaker tags and transcript review still belong in the application.
DeepInfra lists this checkpoint at $0.002 per minute, equivalent to $0.12 for an hour of audio at that listed rate. This is one provider’s price, not a universal tariff. Compare hosted charges or GPU costs using your recording lengths and concurrency, including retries and review.
Why the benchmark headline needs context
NVIDIA reports strong results, but its launch blog and model card give different English Open ASR word-error figures: 5.33% and 4.71%, respectively. The card dates its results to October 12. Without a reconciled evaluation configuration, these should not be presented as one definitive score.
Word error rate measures transcription mistakes. It does not, by itself, tell you whether a meeting summary identified the right decision or attributed a sensitive statement to the right person. For a product evaluation, keep those failure cases separate.
The model card also says its optimized NIM throughput figures use serving optimizations that are not yet in the open-source release. Downloading the checkpoint does not reproduce the fastest advertised serving stack.
Where the release is most useful
The promising use case is a recording workflow with a clear output contract: what wording to preserve, which speaker distinctions matter, and how names and specialist terms should be represented. Start with those requirements and a small representative evaluation set before deciding whether to fine-tune.
NVIDIA also released Granary v2 and adaptation tooling alongside the model. That offers a path beyond prompting, but it adds another decision: whether the remaining errors justify maintaining training data and a custom model.
For speech products, the next useful comparison is the amount of reliable, reviewable work produced per recording. Raw transcription speed is only one part of that answer.
I'm Paras. I work at the intersection of AI, engineering and distribution.
I help AI teams explain technical launches and turn product details into stories developers and founders can use and share.