As of: 2026-11-13 13:49 CET.
Researched and written by AI agents (Claude Opus 5.5 in Claude Code), with a human editor. How we work: https://postcutoff.com/about/
Canonical page: https://postcutoff.com/m/nemotron-3-5-transcribe/

# NVIDIA Nemotron 3.5 Transcribe (30B-A4B)

NVIDIA, Nemotron 3.5 family. Status: current. Type: speech and audio. Released: 2026-10-28.

## How to call it

| Provider | Model id | Endpoint | Docs |
|---|---|---|---|
| Hugging Face (open weights, BF16; FP8 and NVFP4 variants) | `nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B` |  | https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B |
| Hugging Face (optional DFlash2 draft model for speculative decoding) | `nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B-DFlash2` |  | https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B-DFlash2 |
| Self-hosted vLLM 0.32.0 (OpenAI-compatible server) with the NeMo Speech main branch (VLLM_PLUGINS=nemo_speechlm) |  |  | https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B |
| NVIDIA NIM / build.nvidia.com |  |  | https://build.nvidia.com |
| NGC |  |  | https://registry.ngc.nvidia.com/orgs/0767305323357365/teams/transcribe-3_5-ea/models/nvidia-nemotron-3-5-transcribe-30b-a4b |

## Pricing

- Price not published.
- Last checked: 2026-10-29

## Specs

- Input: audio, text
- Output: text
- Context window: 65,536 tokens
- Max output: not published
- Open weights: yes
- Licence: openmdw-1.1

## Where it stands

- No newer speech and audio model in the registry yet.
- Before it from NVIDIA: NVIDIA NemotronLabs VoiceChat 11B (and PersonaPlex-7B) (2026-08-03) https://postcutoff.com/m/nemotron-voicechat/
- Weights are public. Licence: OpenMDW-1.1. See the licence for commercial terms.

## What stands out

- One model for transcription, diarization, translation and audio Q&A: Speech-augmented LLM (ASR + speaker encoders feeding the hybrid Mamba-Transformer MoE Nemotron 3.5 Lightning backbone): verbatim or written-form transcripts, speaker tags, word timestamps (CTC head), translation to and from English, language ID, PII redaction and questions about the recording, chosen by the prompt; 40 languages; recordings up to 1 hour. Source: https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B
- Top open-weights scores on the Open ASR Leaderboard sets (vendor-run): NVIDIA's runs (captured Oct 12, 2026): English Open ASR mean WER 4.71% (Qwen3-ASR-1.7B 4.94), long-form 10.98% (Parakeet-TDT-0.6B-v3 13.37), multilingual Open ASR (7 languages) 4.62% (Qwen3-ASR-1.7B 6.49), multi-speaker mean cpWER 15.80%. Independent leaderboard listing not checked. Source: https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B
- Contextual biasing by prompt: Domain terms and names can be passed in the instruction ('Context terms: ...') to improve recognition of jargon. Source: https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B

## Notes

34B total parameters, ~4.3B active (3.5B language backbone + ~0.8B joint audio encoder). Offline (batch) model, not a streaming ASR service; for real-time use NVIDIA's 0.6B streaming models (see nvidia-parakeet-canary). Runs on one H100 80 GB (BF16) and, quantized to NVFP4, on an RTX 5090; up to 2,238 s of audio per second on one RTX PRO 6000 Blackwell (NVFP4 + DFlash2, NVIDIA). Audio understanding and reasoning were evaluated only in English. Training: ~1.72M unique hours, including Granary v2 (public manifest 463,062 h, 48 languages) and data licensed from David AI.

Quick start (from the model card): `vllm serve ./NVIDIA-Nemotron-3.5-Transcribe-30B-A4B` with `VLLM_PLUGINS=nemo_speechlm`, then send the
audio to the OpenAI-compatible endpoint with an instruction such as "Transcribe with punctuation and capitalization." Keep the system
prompt empty for transcription and translation; use temperature 0 and reasoning off. Technical report:
https://research.nvidia.com/labs/nemotron/files/NVIDIA-Nemotron-3.5-Transcribe-Technical-Report.pdf

Model page: https://postcutoff.com/m/nemotron-3-5-transcribe/