Nemotron Transcribe 3.5

Release Summary

NVIDIA's open-weight speech model for recorded audio. Checkpoint NVIDIA-Nemotron-3.5-Transcribe-30B-A4B: 34B total parameters and about 4.3B active, including the audio encoder, built on the Nemotron 3.5 Lightning 30B hybrid language model. OpenMDW-1.1 weights. Text and audio in, text out. The model card documents a 65,536-token context shared by the prompt and the output, and recordings up to one hour. It transcribes 40 languages, translates 31 of them to or from English, and attributes up to eight speakers in 13 languages. Mean word error rate is 4.71% on the nine public English Open ASR sets, 4.62% on the multilingual Open ASR leaderboard, and 10.98% on the long-form Open ASR sets. Speaker-attributed mean cpWER on full AMI, CALLHOME, Mixer 6, and NOTSOFAR-1 sessions is 15.80%. Preference-ASR word error rate is 7.56% with the default style and 6.16% when the style is instructed. FP8, NVFP4, and GGUF checkpoints ship beside BF16. DFlash2 is a draft model for speculative decoding, not a separate release. The NVFP4 checkpoint with DFlash2, served by the Nemotron Transcribe NIM, transcribes up to 2,238 seconds of audio per second on one RTX PRO 6000 Blackwell GPU. No first-party token price and no measured price per task. Hosted file model in this catalog: Voxtral Small Transcribe 3.

Timeline