NVIDIA Releases Nemotron 3.5 Transcribe: An Open 30B MoE Model for 40-Language Transcription with ~4B Active Parameters

NVIDIA has released Nemotron 3.5 Transcribe, an open-weights speech LLM built for transcription. One model handles 40 languages and recordings up to 1 hour. A text instruction picks the output style for each request. Options include verbatim, written form, speaker-attributed, contextual and translated output. For developers, it can replace separate ASR, diarization and alignment tools.

TL;DR

Size: 34B total and about 4.3B active parameters; 65,536-token context; audio up to 1 hour.
Runs on: 1x H100 80 GB in BF16. FP8 (about 35 GB) and NVFP4 (about 21 GB) checkpoints exist; NVFP4 runs on an RTX 5090. GGUF files target NeMo-Speech.cpp.
Performance: Best open-weights model on the English and multilingual Open ASR Leaderboards; top commercial APIs still lead.
Best: 4.71% mean WER on the 9 public English Open ASR sets (next open model: 4.94%).
Worst: 34.95% cpCER on AliMeeting Mandarin far-field meetings.
Bottom line: Best thing: 1 open model covers transcripts, speaker labels, timestamps and translation. Worst thing: offline only, with weak Mandarin meeting accuracy.

What is Nemotron 3.5 Transcribe?

Nemotron 3.5 Transcribe is a speech-augmented language model (SALM) from NVIDIA that turns audio and instructions into text. An audio encoder feeds embeddings into an LLM decoder, which makes transcription promptable. NVIDIA’s technical research report argues there is no single correct transcript. Numbers, fillers and domain terms depend on who reads the output. NVIDIA positions the model as a transcription specialist, not a general-purpose assistant.

How does the architecture work?

It has three parts:

Parallel Expert Encoder: a trainable 0.6B Transformer ASR encoder runs beside a frozen Nemotron 3 Diarization model. The diarizer tracks up to 8 speakers. Its speaker activity is added to the ASR features every 80 ms. The LLM therefore sees who spoke when, without architecture changes.
LLM backbone: Nemotron 3.5 Lightning 30B-A3B, a hybrid Mamba-Transformer MoE. It has 52 layers: 23 Mamba-2, 23 MoE and only 6 attention layers. Each MoE layer routes to 6 of 128 experts plus 1 shared expert. Few attention layers keep the KV cache small for hour-long audio.
Add-ons: a shared multi-token prediction (MTP) layer drafts tokens for speculative decoding. A 50M-parameter CTC adapter aligns the finished transcript to audio for word timestamps.

The model name says 30B-A4B, but the checkpoint holds about 34B parameters. The extra weight comes from the 1.3B MTP module and the audio components.

Training used 1 supervised fine-tuning stage on 256 H100 GPUs for 28,500 steps. A short RL stage followed: 400 GSPO updates with a reward of 1 minus WER. The model saw about 1.72M unique hours of audio, or 6.5M hours with repeats.

How does it score on benchmarks?

All numbers below are self-reported by NVIDIA in the technical report.

English Open ASR (9 public sets): 4.71% mean WER. The next open model, Qwen3-ASR-1.7B, scores 4.94%. Seven commercial systems still lead, topped by OtterAI speech-v3.1 at 3.88%.
Multilingual Open ASR (7 languages): 4.62% mean WER, versus 6.49% for Qwen3-ASR-1.7B. ElevenLabs Scribe v2 reaches 3.56%.
Long-form ASR: 10.98% mean WER, versus 13.37% for Parakeet-TDT-0.6B-v3.
Multi-speaker, full sessions: 15.80% mean cpWER, versus 16.83% for DiCoW v3.3 with DiariZen. NVIDIA says it is the first SALM to beat cascaded systems here.
Contextual ASR: an entity list cuts named-entity WER on ContextASR-Bench from 12.21% to 2.20%.

<div class="pane on" id="p1">
<div class="flow">
<div class="node on" data-s="0"><span class="n">01</span><b>Audio in</b><span>16 kHz mono, split into 30 s encoder chunks</span></div>
<div class="node" data-s="1"><span class="n">02</span><b>Parallel Expert Encoder</b><span>0.6B ASR encoder + frozen diarizer</span></div>
<div class="node" data-s="2"><span class="n">03</span><b>Hybrid MoE LLM</b><span>Nemotron 3.5 Lightning 30B-A3B</span></div>
<div class="node" data-s="3"><span class="n">04</span><b>Text + timestamps</b><span>CTC adapter aligns words to audio</span></div>
</div>
<div class="wire"><i id="nt35w"></i></div>
<div class="detail" id="nt35d"></div>
<div class="ctrl">
<button class="btn" id="nt35prev">← Back</button>
<button class="btn pri" id="nt35next">Next step →</button>
<button class="btn" id="nt35play"> Auto-play</button>
</div>
</div>

<div class="pane" id="p2">
<div class="seg">
<button class="chip on" data-b="en">English Open ASR</button>
<button class="chip" data-b="ml">Multilingual Open ASR</button>
<button class="chip" data-b="lf">Long-form</button>
</div>
<div id="nt35bars"></div>
<p class="note" id="nt35bn"></p>
</div>

<div class="pane" id="p3">
<p class="note" style="margin:0 0 10px!important">1 checkpoint, different instructions. Click a style to see the request pattern and output shape (illustrative, not a model sample).</p>
<div class="seg" id="nt35ps">
<button class="chip on" data-p="0">Verbatim</button>
<button class="chip" data-p="1">Written form</button>
<button class="chip" data-p="2">Context terms</button>
<button class="chip" data-p="3">Speakers</button>
<button class="chip" data-p="4">Translate</button>
</div>
<div class="out"><div class="lbl">Instruction</div><pre><code id="nt35pi"></code></pre></div>
<div class="out"><div class="lbl">Output shape</div><pre><code id="nt35po"></code></pre></div>
</div>

<div class="pane" id="p4">
<div class="grid2">
<div class="card on"><h4>BF16</h4><div class="big">67.5 GB</div><p>Reference weights · 1x H100 80 GB</p><div class="more">Ampere, Hopper and Blackwell. NVIDIA estimates about 410 RTFx on 1 H100 with MTP2 (batched English ASR).</div></div>
<div class="card"><h4>FP8</h4><div class="big">~35 GB</div><p>Ada Lovelace, Hopper, Blackwell</p><div class="more">Error rate 1.0% higher in relative terms on average. NVIDIA recommends it when non-English accuracy matters most.</div></div>
<div class="card"><h4>NVFP4</h4><div class="big">~21 GB</div><p>Blackwell only · runs on RTX 5090</p><div class="more">Error rate 3.6% higher in relative terms on average. Best when memory or throughput is the main constraint.</div></div>
<div class="card"><h4>NIM speed</h4><div class="big">2,238x</div><p>Real time on 1 RTX PRO 6000 Blackwell</p><div class="more">NVFP4 + DFlash2 in the Nemotron Transcribe NIM; 30 s clip in 0.13 s. These serving optimizations are not open source yet.</div></div>
</div>
<p class="note">Runtimes: vLLM 0.32.0 + NeMo Speech plugin, Hugging Face Transformers, SGLang-Omni 0.2.0, GGML (NeMo-Speech.cpp). License: OpenMDW-1.1.</p>
</div>

<div class="pane" id="p5">
<div class="row"><div class="ic up">+</div><div><b>Best open-weights ASR:</b> 4.71% English and 4.62% multilingual mean WER on the Open ASR Leaderboards.</div></div>
<div class="row"><div class="ic up">+</div><div><b>Meetings:</b> 15.80% mean cpWER on full sessions, ahead of DiCoW v3.3 + DiariZen at 16.83%.</div></div>
<div class="row"><div class="ic up">+</div><div><b>Context:</b> an entity list cuts named-entity WER on ContextASR-Bench from 12.21% to 2.20%.</div></div>
<div class="row"><div class="ic dn">!</div><div><b>Commercial gap:</b> OtterAI speech-v3.1 reaches 3.88% English WER; ElevenLabs Scribe v2 reaches 3.56% multilingual.</div></div>
<div class="row"><div class="ic dn">!</div><div><b>Mandarin meetings:</b> 32.91% cpCER on AISHELL-4 and 34.95% on AliMeeting.</div></div>
<div class="row"><div class="ic dn">!</div><div><b>Offline only:</b> not a streaming ASR service; release-day vLLM is roughly 2x slower than the NIM.</div></div>
</div>

<div class="ft"><span>Source: NVIDIA technical report and model card (self-reported), Oct 2026</span><span class="br">© Marktechpost</span></div>

How do you deploy Nemotron 3.5 Transcribe?

Weights ship in BF16, FP8, NVFP4 and GGUF. A DFlash2 draft model is optional. The OpenMDW-1.1 license allows commercial use.

Runtimes include vLLM 0.32.0 with the NeMo Speech plugin, Transformers and SGLang-Omni 0.2.0. vLLM exposes an OpenAI-compatible API.

NVIDIA model’s main speed comes from its Nemotron Transcribe NIM. With NVFP4 and DFlash2, 1 RTX PRO 6000 Blackwell Server Edition transcribes 2,238 seconds of audio per second. It returns a 30-second clip in 0.13 seconds. There is a catch: those serving optimizations are not open source yet. NVIDIA expects the release-day vLLM path to be roughly 2x slower. On 1 H100 in BF16 with MTP2, NVIDIA estimates about 410 RTFx.

For fine-tuning, NVIDIA tested 17 tasks. Each took about 0.5 to 2.3 GPU-hours on a single GB300.

How does it compare with other open ASR models?

Feature	Nemotron 3.5 Transcribe	Qwen3-ASR-1.7B	Canary-Qwen-2.5B	Whisper-large-v3
Parameters	34B total, ~4.3B active	1.7B	2.5B	1.55B
Languages	40	30 + 22 Chinese dialects	English only	99 (HF tag)
Max audio per request	1 hour	Not disclosed	40 s in training	30 s window, chunked long-form
Speaker attribution	Built in, up to 8 speakers	Not listed	Not listed	Not listed
Word timestamps	Built in (CTC adapter)	Via separate ForcedAligner-0.6B	Not listed	Yes, word-level
License	OpenMDW-1.1	Apache-2.0	CC-BY-4.0	Apache-2.0
English Open ASR, 9 public sets (mean WER, lower is better)	4.71%	4.94%*	5.38%*	6.66%*
Open ASR mean WER on own model card (earlier leaderboard set)	Not reported	5.76%	5.63%	7.44%
Multilingual Open ASR, 7 languages (mean WER)	4.62%	6.49%*	Not applicable (English only)	7.94%*
Reference hardware	1x H100 80 GB (BF16)	Not disclosed	Tested on A6000, A100, RTX 5090	Not disclosed

*Competitor scores on the current 9-set leaderboard come from NVIDIA’s report, citing Hugging Face leaderboard results from October 2026. The leaderboard set changed since the competitors published their own cards, so the 2 WER rows are not interchangeable.

Key Takeaways

One open model handles transcripts, speakers, timestamps and translation across 40 languages.
4.71% English WER is the best open-weights score; commercial APIs still lead.
About 4.3B active parameters; BF16 fits 1 H100, NVFP4 fits an RTX 5090.
An entity list cuts named-entity WER from 12.21% to 2.20%.
Offline only; Mandarin meetings and code-switching remain weak spots.

Check out the technical report, model weights, project page and live demo. Also, feel free to follow us on Twitter and don’t forget to join our 150k+ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well.

Need to partner with us for promoting your GitHub Repo OR Hugging Face Page OR Product Release OR Webinar etc.? Connect with us

FCC Chair Wants To Let Prisons Jam Illegal Cellphones

Microsoft Posts 19% Revenue Growth as Azure Revenue Jumps 48% – Unite.AI

The post NVIDIA Releases Nemotron 3.5 Transcribe: An Open 30B MoE Model for 40-Language Transcription with ~4B Active Parameters appeared first on MarkTechPost.

Credit: Source link