DownloadNVIDIA Nemotron 3.5 Transcribe 30B-A4B is an open-weight speech and audio model designed for multilingual transcription, translation, speaker diarization, and audio understanding. It contains approximately 34B total parameters with 4.3B active, combining a hybrid Mamba-2–Transformer Mixture-of-Experts language backbone with dedicated speech recognition and speaker encoders. The model supports 40 languages, recordings up to one hour, and a 65,536-token context window. Its capabilities include verbatim and formatted transcription, speaker attribution, word-level timestamps, language identification, code-switching, and audio question answering. Configurable reasoning modes enable advanced audio analysis, while native Multi-Token Prediction supports speculative decoding. Available in BF16, FP8, and NVFP4 precision, it integrates with NVIDIA NeMo Speech, vLLM, SGLang-Omni, and Hugging Face Transformers for speech processing applications.
Features
- 34B total parameters with approximately 4.3B activated
- Hybrid Mamba-2–Transformer Mixture-of-Experts architecture
- Speech recognition and speaker diarization encoders
- Multilingual transcription and translation across 40 languages
- Supports recordings up to one hour with 65,536-token context
- Speaker attribution, word-level timestamps, and code-switching
- Audio question answering with configurable reasoning modes
- BF16, FP8, and NVFP4 variants with NeMo Speech and vLLM support
Project Samples
