Nemotron 3.5 Transcribe is an offline speech model on Hugging Face, NGC and build.nvidia.com as of October 28. It has 34B total parameters with about 4.3B active (including the audio encoder), a 65,536-token context and support for recordings up to one hour. Weights come in BF16, FP8 and NVFP4. NVIDIA's reference setup is one H100 80 GB.
Instead of separate models for transcription, diarization and translation, you change the prompt: verbatim or cleaned-up transcript, speaker tags, a list of context terms for product names, translation to or from English, or a question about the recording. The model card publishes the exact prompts. One example: an experimental prompt that transcribes only the speaker nearest the microphone cut word error rate on a test mixing clean speech with far-off meeting chatter to 4.78%, against 27.32% with a standard single-speaker prompt.
NVIDIA's own results put it first among open-weight models on the English Open ASR Leaderboard (4.71% mean WER against 4.94% for Qwen3-ASR-1.7B) and on long-form audio (10.98% against 13.37%). It isn't a streaming model: the card says plainly it's not for real-time ASR. NVIDIA also says the fast serving path behind its throughput figures isn't in the open-source release yet, and the vLLM integration that shipped on day one is expected to be about two times slower.