Nemotron 3.5 Transcribe
NVIDIA's open ASR model for 40 languages with 5.33% average English WER, released with the 463,062-hour Granary v2 dataset.
Overview
NVIDIA published Nemotron 3.5 Transcribe and the Granary v2 dataset on October 28, 2026. The mixture-of-experts model has about 30B total and 4B active parameters (the Hugging Face listing shows 34B), transcribes 40 languages, translates 31 to or from English, attributes speakers in 13 and handles recordings up to an hour with up to eight speakers. NVIDIA reports 5.33% average WER on the Hugging Face Open ASR English leaderboard (first among open-weight models), 2.8% AA-WER and 18.26% DER on a 21-language diarization test. Granary v2 spans 463,062 hours in 48 languages. Weights are on Hugging Face and ModelScope, with hosted endpoints on Baseten, DeepInfra and Simplismart and self-hosting via vLLM and SGLang-Omni.
This is a sourced directory entry. Features and capabilities are reported in the linked sources; this entry does not document independent Aymorphic testing.
Key features
- 40-language transcription, 31-language translation to or from English, speaker attribution in 13
- 5.33% average English WER on Open ASR, first among open-weight models (NVIDIA-reported)
- Long-form audio up to an hour, up to eight speakers
- About 100 hours of audio in under three minutes on one RTX PRO 6000 Blackwell with an unreleased NVFP4 version
- Granary v2: 463,062 hours across 48 languages, manifests on Hugging Face
Use cases
- Multilingual meeting and call transcription with speaker labels
- Training or adapting speech models on Granary v2 data
Pricing
Open weights on Hugging Face and ModelScope; license not stated in the announcement; hosted endpoint prices unknown.
Pricing reflects the source record and may change. Confirm current terms with the provider.
Dates and sources
- Released
- 28 Oct 2026
- Record updated
- 29 Oct 2026
- Record last checked
- 29 Oct 2026