NVIDIA releases Nemotron 3.5 Transcribe: open weights for recorded audio, with technical detail and verification still pending

In an autumn when voice technology is drawing wide market attention, as shown by recent releases from competitors such as Mistral and ElevenLabs cited by TechCrunch, NVIDIA published Nemotron 3.5 Transcribe on 28 October 2026. Unlike Nemotron 3.5 ASR, the 0.6-billion-parameter model for real-time streaming released in June 2026, the new open-weights release is aimed at analysing audio files that have already been recorded. According to the company's official channels on Hugging Face, it brings transcription, translation, speaker attribution and questions about the speech signal together in a single solution, built on a hybrid Mamba-2/Transformer mixture-of-experts architecture paired with a parallel audio encoder.

On the technical side, the model card describes a total size of about 34 billion parameters (about 4.3 billion active, encoder included), whereas the official name "30B-A4B" and the blog point to about 30 billion total and 4 billion active: the difference depends on whether the audio encoder is counted. Released under the OpenMDW-1.1 licence for commercial and non-commercial use, the model supports Italian in the performance tier the company labels 'Leading'. The recommended hardware is an 80 GB H100 GPU. Alongside the model, NVIDIA made available the open Granary v2 dataset, with 463,062 hours of speech across 48 languages; the original audio files must be downloaded directly from their source platforms, in line with each corpus's licence. Among the use cases cited, adapting the model to the medical domain with the Eka Care benchmark showed that, after about 14 GPU-hours of training on an NVIDIA GB300, WER on clips held out from training fell from 22.44% to 12.17%, a 45.8% relative reduction across 2,764 clips.

A closer look at the published figures, however, reveals some discrepancies and limits to verification. The performance measurements come from internal tests or tests run with partners, with no independent replication, as NVIDIA's own blog admits. There is also a gap between the 5.33% average error rate stated in the post for the English Open ASR leaderboard and the 4.71% given in the model card as an equal-weight average over nine datasets: they are probably two different calculations, but the difference is not explained. The speed figure (100 hours of audio transcribed in under three minutes) likewise refers to an NVFP4 version that has not yet been released. NVIDIA was explicit about the scope, stressing in the blog that "Transcribe is built for recorded audio, not live streaming" and repeating in the official model card that chunked internal processing does not turn the system into a real-time transcription service; in the same way, the labels distinguish up to eight speakers per file without identifying them.

Releasing open weights and datasets gives developers concrete resources, but when a model's effectiveness is measured only by its maker's own metrics, technical transparency stays halfway between the announcement and real verification.

— Olya