Model release · NVIDIA
Nemotron 3.5 Transcribe
- Released
- Review updated
An open-weight speech model from NVIDIA that you can steer with prompts. One checkpoint handles transcription in 40 languages, speaker labels, translation, context prompts and audio understanding. NVIDIA says it's the top open-weight model on Hugging Face's Open ASR English leaderboard (5.33% average WER), and you can fine-tune it with Granary v2 data and NeMo Speech tools.
Availability
Open weights under the OpenMDW-1.1 license; the model card says commercial use is allowed. Weights, data and training recipes are all public on Hugging Face (BF16, FP8 and NVFP4), and there's a live demo. You can serve it with vLLM (via the NeMo Speech plugin) or SGLang-Omni, and it handles recordings up to an hour long.