NVIDIA releases Nemotron 3.5 Transcribe with 40-language support

NVIDIA released Nemotron 3.5 Transcribe, an open-weight speech recognition model, on October 28, 2026. The model weights are published on Hugging Face under the OpenMDW-1.1 license, which NVIDIA lists as a commercial-use license.

Nemotron 3.5 Transcribe uses a mixture-of-experts architecture with about 34 billion total parameters and about 4.3 billion active parameters. It can process recordings up to one hour long and distinguish up to eight speakers, noting when each one speaks.

Users can set the transcription style in natural language, for example verbatim output or cleaned-up sentences. The model recognizes speech in 40 languages and translates among 31 languages, mainly into English.

Benchmark results

NVIDIA says the model ranked first among open-weight models on Hugging Face's English open ASR leaderboard with a word error rate of 4.71 percent. On a seven-language multilingual evaluation it recorded an average error rate of 4.62 percent, and on ContextASR-Bench its context-based recognition cut proper-noun errors to as little as one sixth.

In a medical speech recognition experiment with additional training on proprietary data, word error rate fell 45.8 percent relative to an optimized prompt baseline. NVIDIA also provides the NeMo Speech tool and the Granary 2 speech training dataset.

According to aitimes.com, NVIDIA says one Blackwell-based RTX PRO 6000 GPU can process up to 2238 seconds of audio per second in an optimized inference environment, though not all of those optimizations are included in the public open-source version. The model takes recorded audio as input and does not handle real-time streaming audio directly.