What happened
NVIDIA published Nemotron 3.5 Transcribe on October 28, 2026, making the checkpoint available through Hugging Face and ModelScope. The model is designed for recorded audio rather than live streaming. Its stated scope combines several tasks that are often assembled from separate services: speech transcription, diarization or speaker attribution, translation, and question answering over audio. NVIDIA says it transcribes 40 languages, translates 31 languages to or from English, and attributes speakers in 13 languages. It accepts recordings of up to one hour and can label up to eight speakers.
The release is promptable. A developer can ask the same model for a transcript with speaker labels, an English translation, or an answer about decisions made in a meeting. The source gives a healthcare example in which the instruction supplies possible terms, asks for standard medical spelling and conventional number formatting, and says those terms should only be used when supported by the audio. The model card includes presets, a prompt guide, and a task-by-language matrix. Word-level timestamps are supported, but NVIDIA explicitly advises users to check names, speaker changes, and answers against the recording. Audio Q&A is trained in English.
NVIDIA describes the model as a hybrid Mamba-Transformer mixture-of-experts system with 30 billion total parameters and roughly 4 billion active per token. Its reported results include a 5.33% average word error rate on the English Hugging Face Open ASR Leaderboard among open-weight models, 4.62% mean WER across seven languages, and 10.98% mean WER on the cited long-form comparison. On Artificial Analysis's index, NVIDIA reports a 2.8% AA-WER tie for first among open-weight models. For diarization, it reports an 18.26% diarization error rate across 21 languages on David AI's benchmark and a 32.4% cpWER result on VoiceArena's English full-session benchmark. These are benchmark claims reported by NVIDIA, not a guarantee for a particular recording set.
The company also reports a throughput test: 100 hours of audio in under three minutes on one RTX PRO 6000 Blackwell Server Edition GPU, using a forthcoming NVFP4 quantized version, 30-second clips, and high concurrency. The available checkpoint can be served with vLLM. Quantized versions, local deployment through NeMo-Speech.cpp, and support for RTX Spark, DGX Spark, and Jetson AGX Thor are described as forthcoming.
Alongside the model, NVIDIA released Granary v2, a public manifest and reusable data-curation pipeline built with NeMo Curator. The release covers 463,062 hours across 48 languages and 16 source corpora, compared with 25 languages in Granary v1. NVIDIA cautions that those figures describe the dataset release rather than the full Transcribe training mixture or its sampling weights. The pipeline produces aligned transcript variants, task labels, quality metadata, audio captions, code-switching restoration, speech-grounded Q&A, contextual ASR biasing terms, and translations with automatic quality scores. Its four stages cover metadata preparation, ASR, text-label generation, and high-fidelity filtering. Developers can run it on their own recordings or authorized public corpora using Ray and NVIDIA Dynamo backends.
What the archive adds
This is a new product and data release within NVIDIA's wider Nemotron 3.5 line, rather than a restatement of the earlier multimodal-model announcement. Nemotron 3.5 Super VL concerned an open 120B MoE multimodal model for specialized agent workflows; Transcribe instead targets the speech-input layer and supplies its associated curation pipeline.
Trend signal
No supplied pattern is a close fit. The release provides speech capabilities and adaptation assets, but it does not establish a change in agent protocols, runtime security, or another documented New Runtime pattern.
Why now
The source frames Granary v2 as the data and curation foundation behind the model's promptable output formats. That makes the paired release timely: a team that finds prompting insufficient for its terminology, accents, or transcript conventions is also given the artifacts NVIDIA says can support fine-tuning and data preparation.
Connections to other events
An earlier archive item covered Gemini 3.5 Transcribe as a specialized transcription model and API release. NVIDIA's new release differs in its emphasis on open weights, self-directed processing location, and the Granary v2 manifests and pipeline. Another earlier item distinguished file transcription from live microphone and call workflows; NVIDIA makes the same boundary explicit, stating that Transcribe is for recorded audio, not live streaming.
Why it matters
Speech workflows often fail at the parts surrounding plain text: who spoke, when they spoke, how a domain term should be rendered, whether a translation is needed, and whether a transcript can become a usable meeting artifact. Nemotron 3.5 Transcribe offers a single checkpoint for those outputs and exposes prompt-level formatting control before a team commits to training. Granary v2 adds a more consequential option for organizations with authorized recordings: reproduce a curation path, inspect quality metadata, and adapt to local vocabulary or conventions. NVIDIA reports that fine-tuning in a medical case reduced WER by 45.8% relative to the best prompt on held-out speakers. That is a useful direction, not evidence that every domain will receive comparable gains.
What happens next
- NVIDIA releases the announced quantized and local-deployment variants. (3 months, 70%). Watch for: NVFP4 checkpoint and NeMo-Speech.cpp availability.
- Teams evaluate domain adaptation against prompting for specialized recorded-audio workloads. (6 months, 65%). Watch for: published held-out evaluations by language, domain, and speaker mix.
Limits and uncertainty
The source is an NVIDIA release post, so its leaderboard positions, throughput, and case-study improvement are vendor-presented claims. Several results depend on specific benchmarks, test-set definitions, hardware, quantization, clip length, and concurrency. The post does not establish accuracy for every supported language, accent, medical vocabulary, noisy recording condition, or speaker configuration. It also does not provide a live-streaming solution. Granary v2's 463,062-hour figure is for the public dataset release, not a disclosure of the complete training mixture. Builders must also determine whether their recordings are authorized for the proposed curation and fine-tuning workflow.
Skill to build now
Create a small evaluation set of authorized recorded calls or meetings with verified transcripts, speaker turns, and the domain terms that matter most. Run a prompt-only baseline with explicit formatting and vocabulary context, then score word errors, speaker changes, timestamps, and critical-term accuracy separately. If the baseline misses business-critical terms, use Granary-style transcript variants and quality checks to define what a fine-tuning dataset would need before training.
Reader takeaway
Nemotron 3.5 Transcribe is not just another transcription endpoint: NVIDIA is releasing an open, promptable recorded-audio model together with a large data manifest and reusable preparation pipeline. Its practical value will depend less on headline leaderboard placement than on whether a team can validate its own speakers, terminology, languages, and privacy constraints.
Open question
Can the public Granary v2 pipeline make adaptation economical for smaller domain datasets, or will the compute and data-governance burden keep the strongest gains limited to larger operators?