On this page
NVIDIA released open weights for Nemotron 3.5 Transcribe on October 28, 2026. The main Hugging Face checkpoint is nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B. It is an offline speech and audio model, not a real-time streaming ASR service. Related open Nemotron work on this site (vision/Bolt, not Transcribe): Nemotron 3.5 Super VL and Bolt: Specs, License, Runs on 24GB.
Size and capabilities
The NVIDIA AI account describes a promptable 30B MoE with about 4B active parameters. The Hugging Face model card table lists 34B total (4.3B active) including the audio encoder: roughly a 3.5B-active language backbone plus about 800M for the joint audio encoder. The product ID / social wording and the card table are layered figures from different sources, not one merged number.
Per the card and NVIDIA AI posts, the model covers transcription in about 40 languages, plus speaker-attributed transcription, speech translation, contextual transcription, and audio question answering. Recordings can run up to about 1 hour, with a 65,536-token context. License on the card: OpenMDW-1.1 (commercial or non-commercial use as stated there).
Open ASR and AA-WER figures (separate sources)
In a follow-up post, the NVIDIA AI account says the model ranks #1 among open-weight models on the Hugging Face Open ASR English leaderboard with a 5.33% average word error rate. That is NVIDIA AI’s own citation. Separately, the Hugging Face card’s English Open ASR Leaderboard table (nine public sets, equal-weight mean; results captured October 12, 2026) shows 4.71% WER. The 5.33% and 4.71% figures differ in source and timing, so they are not one “official” score.
Artificial Analysis posted 2.8% on its AA-WER Index, calling it a shared open-weight #1 with Mistral Voxtral Small. That is a third-party index figure, not an NVIDIA official result and not retested here.
For other speech-model context on this site: Block Releases Dessa, a Full-Duplex Speech Model Trained Only on Speech and Which IBM Granite Speech Model Fits Your Recording?.
Weights and task details follow the Hugging Face model card and the cited NVIDIA AI / Artificial Analysis posts. For reference only.