Nemotron 3.5 Transcribe: Specs, License & How to Run

Nemotron 3.5 Transcribe is NVIDIA’s open-weight offline speech model: a 30B-A4B Speech-Augmented Language Model. Official hosts we opened on 2026-10-29: the ungated card nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B (created 2026-10-07T22:50:21Z; last modified 2026-10-28T21:01:32Z; 182 likes / 338 downloads), the OpenMDW-1.1 LICENSE, the optional draft nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B-DFlash2 (6 likes / 37 downloads; ungated), and the card’s clone URL github.com/NVIDIA-NeMo/Speech. Card release line: Hugging Face 10/28/2026. API gated was false. Hugging Face pipeline_tag is automatic-speech-recognition.

This page hangs off Otter.ai alternatives and Descript alternatives. Those mappings are editorial. Otter is a hosted meeting agent. Descript is a timeline editor. This card transcribes or translates a recording. The card says it is not a real-time streaming ASR service. We do not claim product parity.

Same Nemotron 3.5 family, different jobs: Nemotron 3.5 Bolt (74B/9B local-agent VLM) and Nemotron 3.5 Super VL (120B/12B VLM). Do not treat Transcribe as a chat checkpoint. Same-site file ASR peer: Granite Speech 5.0 — Nemotron 3.5 Transcribe vs Granite Speech 5.0. Streaming speech-to-speech on this site is Hy-SimulMT and Hibiki-Zero, not this card.

What the card says it does

Copied labels from the README we opened: multilingual transcription (verbatim or written form), contextual transcription, speaker-attributed transcription, rich transcription, speech translation, language identification, and audio question answering. Recordings up to one hour. 40 languages on Hugging Face cardData.language. Speaker labels “distinguish speakers within a recording; they do not establish real-world identity.” Internal audio chunking (30-second encoder chunks) “does not make it a real-time streaming ASR service.” Those sentences are upstream.

What we verified (2026-10-29)

Field	Value	Source
Host	huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B (not gated)	HF API
Created / last modified	2026-10-07T22:50:21Z / 2026-10-28T21:01:32Z	same
Likes / downloads	182 / 338	same
License (card)	other / openmdw-1.1	HF cardData
LICENSE file	OpenMDW License Agreement, version 1.1	LICENSE
Pipeline tag	automatic-speech-recognition	HF API
Params (NVIDIA)	34B total / 4.3B active, including audio encoder	HF card “Model Summary”
Card name	30B-A4B	same
safetensors.parameters	33,697,781,029 (~33.7B)	HF API
usedStorage	67,532,739,306 bytes (~62.9 GiB)	same
Architecture	nemo_speechlm / Nemotron3_5TranscribeForConditionalGeneration	config.json
Context (card / vLLM example)	65,536 tokens; recordings up to 1 hour	card + vllm serve --max-model-len 65536
config.json text_config.max_position_embeddings	262,144	config.json (Lightning backbone; serving example still uses 65536)
Base (card)	nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16	card
DFlash2 sibling	ungated; 6 likes / 37 downloads	HF API
FP8 / NVFP4 sibling API	HTTP 401 (not opened)	HF API 2026-10-29

We keep NVIDIA’s 34B / 4.3B quotes. The 30B-A4B filename is the card name. runsLocally: true means the BF16 weights are downloadable; the vendor quick start is one H100 80 GB, not a laptop checkpoint.

How to run (from the official card, not yet tested)

The card’s vLLM path (vLLM 0.32.0, NeMo Speech main, one H100 80 GB):

export VLLM_PLUGINS=nemo_speechlm
export VLLM_MAX_AUDIO_DECODE_DURATION_S=3600
vllm serve ./NVIDIA-Nemotron-3.5-Transcribe-30B-A4B \
  --served-model-name nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B \
  --trust-remote-code --dtype bfloat16 --gpu-memory-utilization 0.94 \
  --max-model-len 65536 --reasoning-parser nemotron_v3

Transformers in-process path on the same README: AutoModelForMultimodalLM / AutoProcessor from nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B. Card sampling: transcription/translation/language-id temperature 0, reasoning off. forkalt has not run these commands.

Clone URL copied from the card: git clone https://github.com/NVIDIA-NeMo/Speech.git nemo-speech. The NeMo Speech toolkit license is a different file from the weight grant. Do not flatten them.

License

LICENSE is titled OpenMDW License Agreement, version 1.1 (OpenMDW-1.1), copyright NVIDIA CORPORATION & AFFILIATES, 2026. Hugging Face license_name is openmdw-1.1. Same SPDX treatment as Nemotron 3.5 Bolt and Nemotron 3.5 Super VL.

Paraphrase only, not legal advice:

• Include the copyright notice and this license with copies or substantial portions.
• Commercial use is described as permitted on this card (“ready for commercial or non-commercial use”).
• Read the file for attribution and Model-as-a-Service terms (the Bolt/Super VL siblings require prominent “Built with NVIDIA Nemotron” attribution).

What we did not verify

Vendor Open ASR Leaderboard / long-form / multilingual / multi-speaker cpWER rows stay self-reported (card: results captured 12 October 2026). The RTX PRO 6000 “2,238 seconds of audio per second” NIM figure is vendor and the card says those serving optimizations are not yet in the open-source release. Peak VRAM is not independently measured. We have no sourced bake-off against Otter, Descript, Granite Speech 5.0, whisper.cpp, or pyannote community-1.

Related: Granite Speech 5.0 — Nemotron 3.5 Transcribe vs Granite Speech 5.0, Hy-SimulMT (streaming zh–en S2ST, not file ASR), Hibiki-Zero, pyannote community-1 (diarization-only), whisper.cpp, vLLM, Nemotron 3.5 Bolt and Nemotron 3.5 Super VL (VLMs, not this card). Catalog: home. Editorial: About.