NVIDIA releases Nemotron 3.5 Transcribe and Granary v2

NVIDIA released Nemotron 3.5 Transcribe and Granary v2, bringing together a downloadable multilingual speech model, public data manifests and curation tools for adapting speech systems.

The model joins the same product family as Nemotron 3.5 Super VL and Bolt, but it is built for recorded speech. The published checkpoint handles transcription, speech translation, speaker-attributed transcripts, word-level timestamps, language identification and audio question answering. It is not designed for live streaming.

NVIDIA’s support matrix lists transcription in 40 languages, translation between English and 31 of those languages, and multi-speaker transcription in 13 languages. Inputs are limited to one hour and eight speakers, and audio reasoning was evaluated only in English.

The model card and technical report describe a 34-billion-parameter system that activates about 4.3 billion parameters per request, including its audio encoders. It uses a 30-billion-parameter hybrid Mamba-Transformer mixture-of-experts language backbone and is released under the OpenMDW 1.1 license.

Granary v2 combines public manifests with a reusable curation and labeling pipeline. NVIDIA says the public release covers 463,062 hours across 48 languages and 16 source corpora, with three transcript forms plus task labels, translations and quality metadata. The package does not include all underlying audio; users must obtain the source recordings separately and comply with each corpus’s license.

Those figures do not represent the model’s complete training set. NVIDIA’s technical report says the released manifests contain only the shareable portion curated with the pipeline. The full training mixture also includes private multilingual speech and other task-specific sources.

Independent benchmark evidence is narrower than NVIDIA’s performance claims. Nemotron 3.5 Transcribe ranks second among open-weight speech-to-text models at 2.8% on its AA-WER measure, tied on the displayed score with the first-listed model. The evaluation weights results by audio duration across roughly eight hours from three datasets.

NVIDIA’s own English Open ASR figures conflict. Its launch article reports a 5.33% average word error rate; the model card and technical report report 4.71%. The underlying leaderboard did not expose enough rendered data to resolve the discrepancy.

NVIDIA also says an optimized deployment processed as much as 2,238 seconds of audio per second on one RTX PRO 6000 Blackwell Server Edition GPU. The reviewed sources did not independently reproduce the result, and the model card says the serving optimizations used in the test were not yet included in the open-source release. NVIDIA warns that outputs can include substitutions, deletions, repetitions, hallucinated speech, incorrect language selection and speaker-attribution errors.