Nvidia released Nemotron 3.5 Transcribe, an open-weight 30B speech model

A product team that needs to turn an hour of meeting audio into text has two routes today. Pay a cloud service, or stand up a model on their own servers. Nvidia on 28 October released Nemotron 3.5 Transcribe, a 30 billion parameter open-weight speech model that fits the second route. The model handles 40 languages and reads recordings up to one hour long. It is on Hugging Face, GitHub and Nvidia's own build site under the OpenMDW-1.1 licence.

What it does

The model card lists 40 supported languages, grouped by quality level. English, French, German, Spanish, Hindi, Mandarin, Japanese, Korean and Arabic are in the top group. Nvidia says the model can also translate 31 of those languages to and from English. On the Open ASR Leaderboard, Nvidia reports a mean word error rate of 4.71 percent on English. The multilingual mean is 4.62 percent. On long-form audio the mean rate is 10.98 percent. On multi-speaker audio the character-weighted rate is 15.80 percent.

The architecture is a hybrid of Mamba-2 and Transformer blocks with a mixture-of-experts design. The card gives the model as 34 billion total parameters with 4.3 billion active per step. Speaker diarization is built in, tagged <spk:0>, <spk:1> and so on. The card says the model was tested with up to eight speakers in constructed mixtures.

Only recordings, not live audio

The card calls this out directly. It says: "This is an offline model. Internal audio chunking does not make it a real-time streaming ASR service." TechCrunch describes the limit the same way. It writes that the model "only works on recordings, not live audio". That rules out live captioning and real-time meeting assistants.

Nvidia's own Parakeet 1.1B is the streaming option for anyone who needs real-time. The new model fits meeting notes written after the call. It also fits medical transcription on recorded consultations and batch work on a backlog of audio.

Hardware

The model needs at least one H100 80 GB or a newer Hopper or Blackwell Nvidia GPU. Nvidia lists Jetson AGX Thor, DGX Spark and Blackwell datacenter chips as supported. Quantised variants in FP8 and NVFP4 are published next to the BF16 reference weights.

Context length is 65,536 tokens. The card says the model is not guaranteed to work beyond one hour of audio.

Who can use it

The OpenMDW-1.1 licence is the Open Model Derivatives and Weights licence. Nvidia published the model on Build.Nvidia.com, GitHub's NeMo repository, Hugging Face and NGC on the same day. Weights, inference code and the quantised variants are available.

Source

Nemotron 3.5 Transcribe on Hugging Face. Reporting by TechCrunch and The New Stack.