Back to news
launchNVIDIA2026-10-28

NVIDIA open-sources Nemotron 3.5 Transcribe supporting 40 languages and a 65K context

NVIDIA released the open-weight Nemotron 3.5 Transcribe speech model on October 28, with BF16, FP8 and NVFP4 weights published under the OpenMDW-1.1 license for commercial and non-commercial use. The model uses a hybrid Mamba-2 and Transformer mixture-of-experts architecture with about 34 billion total parameters, roughly 4.3 billion active including an approximately 800 million parameter audio encoder, and a context length of 65,536 tokens. It handles recordings up to one hour, distinguishes up to eight speakers, transcribes 40 languages and translates between 31 languages and English. NVIDIA reports average word error rates of 4.71 percent on English, 10.98 percent on long-form and 4.62 percent on multilingual benchmarks, all runnable on a single H100 80GB GPU, and published the model card, technical report and serving documentation.

NVIDIANemotron 3.5 Transcribe开源语音转写多语言OpenMDW