NVIDIA releases open-weight Nemotron 3.5 Transcribe speech model and Granary v2 pipeline

NVIDIA released NVIDIA-Nemotron-3.5-Transcribe-30B-A4B on Hugging Face on 28 October 2026 under the permissive OpenMDW-1.1 licence, alongside its Granary v2 data curation pipeline and NeMo fine-tuning recipes. Built as an offline Speech-Augmented Language Model, the checkpoint pairs a speech encoder with a 34-billion-parameter hybrid mixture-of-experts decoder that activates roughly 4.3 billion parameters per token and ships in BF16, FP8, and NVFP4 quantised weights.

Within a 65,536-token context window, the model ingests continuous recordings up to one hour long without external chunking. A single unified checkpoint handles automatic speech recognition across 40 languages, bidirectional speech-to-text translation between English and 31 languages, speaker diarisation across 13 languages, and direct question answering over raw audio.

Key numbers

• 5.33% WER: #1 open-weight score on the Hugging Face Open ASR English Leaderboard.
• 2.8% AA-WER / 1.8% AA-AgentTalk WER: Tied #1 open-weight accuracy on Artificial Analysis Speech-to-Text.
• 34B total / ~4.3B active parameters: Hybrid MoE weights available in BF16, FP8, and NVFP4 formats.