For every token of the recording, the model lights up about 4 of its 30 billion parameters.
Fact
The facts: on October 28, 2026 NVIDIA published Nemotron 3.5 Transcribe on its Hugging Face blog - an open-weight model you can prompt. Only about 4B of its 30B parameters are active per token. The model transcribes 40 languages, translates 31 of them to or from English and attributes speakers in 13. A second NVIDIA post from the same day says the model is available on Hugging Face, with quantized versions coming soon. Alongside the model comes Granary v2: its public manifest covers 463,062 hours across 48 languages and 16 source corpora; the first version had 25 languages. According to NVIDIA the model is first among open-weight models on the Open ASR Leaderboard at 5.33% average word error rate (English), and on Artificial Analysis it scores 2.8% AA-WER and shares first place among open models. In a medical example fine-tuning cuts the error from 22.44% to 12.17%. With the upcoming NVFP4 version on a single RTX PRO 6000, NVIDIA promises 100 hours of audio in under three minutes.
Word error rate is the measure behind these numbers: the percentage of words the model writes wrong. The numbers come from NVIDIA's announcement. The Open ASR Leaderboard and Artificial Analysis are outside rankings, but NVIDIA is the one citing them, and the medical example and the speed are its own.
Granary v2 is the other half of the news. The first version had 25 languages, the second reaches 48, and the manifest is public. What exactly you can download beyond the manifest, our source does not say.
One number does not fit. The model card on Hugging Face shows 34B including the audio encoder, while NVIDIA's text says 30B. Here I follow the text.
Bulgarian is among the 40 languages: the model card lists it for transcription and translation to and from English, without multispeaker ASR. The license is OpenMDW-1.1. The date of the quantized versions is missing. Open the model card before you build on it.
Official primary sources
NVIDIA on Hugging Face - Nemotron 3.5 Transcribe and Granary v2, 28.10.2026NVIDIA blog - local AI on RTX Spark, 28.10.2026NVIDIA - Nemotron 3.5 Transcribe model card on Hugging FaceThe visual is generated code-art, no third-party images.