NVIDIA Ships Nemotron 3.5 Transcribe: 40 Languages, Diarization and Translation in One Open Model
NVIDIA released Nemotron 3.5 Transcribe, an open-weight speech model that transcribes 40 languages, translates 31 and separates up to eight speakers in an hour-long recording. The interesting part is not the size: it is that you drive it with plain-language instructions instead of configuration.
What NVIDIA Shipped
The announcement landed Wednesday, October 28. The model is available on Hugging Face and ModelScope, alongside Granary v2, the speech data pipeline it was trained on, plus fine-tuning recipes.
A 30B Model With 4B Active
Transcribe carries 30 billion parameters, but only about 4 billion activate per token: it pairs a speech encoder with a hybrid Mamba-Transformer mixture-of-experts language model. That is what lets it fit on a single chip, from Jetson AGX Thor up to a Blackwell data-center GPU.
Granary v2: 463,000 Hours of Audio
The public manifest covers 463,062 hours across 48 languages and 16 source corpora, up from 25 languages in the previous version. It includes three transcript forms — raw, punctuated and normalized —, per-record quality flags and the curation pipeline built with NeMo Curator.
Read also
The Numbers That Matter
Open ASR Leaderboard and AA-WER
On Hugging Face's leaderboard it ranks first among open-weight models in English at 5.33 percent average error, and in multilingual at 4.62 percent across seven languages. On Artificial Analysis' AA-WER index it ties for first at 2.8 percent using 4B active parameters against the 24B model it ties with.
Diarization: Who Spoke and When
It separates speakers in 13 languages and marks word-level timestamps. NVIDIA says it posts the lowest error on David AI's 21-language diarization table and ranks first in VoiceArena's 14-system comparison. Those are the vendor's numbers: you read them, you don't re-run them.
What Changes for a Real App
Instructions Instead of Configuration
The same checkpoint answers a written prompt: "transcribe and label each speaker", "translate this audio into English" or "what decisions were made in this meeting". For anyone integrating it, that replaces a chain of separate models with one call.
Fine-Tuning: 45.8 Percent Fewer Errors in a Medical Case
NVIDIA reports that, in a medical dictation case, fine-tuning with its recipes cut error on held-out speakers by 45.8 percent relative to the best prompt. That is the most honest argument in the announcement: the base model works, but the value sits in adapting it to domain vocabulary.
The Limits, Stated Plainly
It Does Not Do Live Audio
It is built for recordings, not streaming. If your product needs real-time captions, this model is not the part. There is also no published quantized version yet and no local binary: NVIDIA says those are coming.
Where It Runs and What Is Missing
Today it serves through vLLM. The company measured 100 hours of audio in under three minutes on one RTX PRO 6000 Blackwell, but with the NVFP4 version that is not out yet and on 30-second clips at high concurrency: useful as a ceiling reference, not a promise for your workload.
Conclusion
Nemotron 3.5 Transcribe moves open transcription from "a model" to "a product piece": one checkpoint with diarization, translation and questions about the audio. If you come from running models on your own machine, the local AI guide with Ollama is still the place to start without paying for an API.
