What happened
NVIDIA released Nemotron 3.5 Transcribe on October 28, 2026. TechCrunch describes it as a free, open-weight model that can transcribe audio in up to 40 languages, translate 31 languages to and from English, process recordings up to an hour long, and distinguish as many as eight speakers. NVIDIA also released version 2 of Granary, its open-source speech dataset.
The supplied technical report describes a model with about 34 billion parameters in total and 4.3 billion active parameters. Its components include a Transformer automatic-speech-recognition encoder, a frozen Nemotron 3 Diarization model, and the Nemotron 3.5 Lightning 30B-A3B hybrid Mamba–Transformer mixture-of-experts language model. The report names the released checkpoint NVIDIA-Nemotron-3.5-Transcribe-30B-A4B and lists BF16, FP8, NVFP4 and GGUF formats. The model is released under the OpenMDW-1.1 license.
The report says the system supports verbatim, written-form, contextual, speaker-attributed and rich transcription, along with speech translation and spoken-language identification in 40 languages. It can produce speaker-attributed word-level timestamps. That range matters because developers need not treat transcription, speaker labeling and translation as separate model calls for every workflow; the source describes a single model supporting those output modes. The supplied material calls it a speech-augmented language model, which frames the release as more than a conventional speech-to-text checkpoint.
Granary v2 is the accompanying dataset and curation pipeline. According to the supplied release material, it expands language coverage from 25 to 48 languages, with public labels and instruction pairs. NVIDIA says developers can combine the dataset and NeMo Speech tooling to customize applications for industry jargon in different languages. The public labels and instruction pairs are described as released on Hugging Face, while the model card is said to list distribution through Hugging Face, GitHub, NGC and build.nvidia.com.
TechCrunch reports that NVIDIA positions the release for business meetings and medical transcription, and says the model can run on a single chip, ranging from Jetson AGX Thor and DGX Spark to Blackwell data-center chips. This is an availability claim rather than a measured deployment comparison: the supplied material does not state which checkpoint format, audio duration, hardware configuration, throughput, memory use, or latency applies to each environment. The stated functional limitation is clearer: Transcribe is for recorded audio, not livestreams. NVIDIA's claim that it beats competitors on benchmarks is reported, but the supplied material does not include benchmark results.
What the archive adds
This is a restatement of the same release, not a new event or new fact on a known story. The archive already covered NVIDIA Nemotron 3.5 Transcribe and Granary v2 and the open-weight promptable speech model. Those notes establish that multilingual transcription, diarization, translation, audio Q&A, the data pipeline, and fine-tuning recipes were already part of the published account.
Trend signal
This release is consistent with generative media becoming application infrastructure: the package joins a model, a dataset and customization tooling for developers to incorporate audio capabilities into applications. Its described outputs—transcripts, translation, speaker attribution and timestamps—are components an application can consume rather than a one-off generated artifact. The supplied material does not establish how reliably those components perform in production, so the signal is about the direction of the product boundary, not proven deployment outcomes.
Why now
The release arrives amid a busy period for voice AI. TechCrunch notes recent launches from Mistral, ElevenLabs and Salesforce, making NVIDIA's announcement part of a broader wave of speech products rather than an isolated research publication. NVIDIA's specific combination is open weights, a companion dataset and adaptation tooling. The source does not explain a more specific trigger for the timing, nor does it establish whether the competing releases have equivalent capabilities.
Why it matters
For teams building meeting, support or medical-transcription products, a single open-weight model that supports multilingual transcription, translation, speaker attribution and word-level timestamps could simplify the speech pipeline. Instead of assembling a separate transcription service, diarization stage and translation layer, a team can evaluate whether the described unified system fits its recorded-audio workflow. That does not mean the tasks are necessarily equally strong, but it makes the integration boundary worth testing.
The accompanying dataset and NeMo Speech tooling also give developers a stated route to adapting recognition to domain vocabulary, rather than relying only on a general-purpose transcript. This is especially relevant where names, abbreviations and specialized terminology affect the usefulness of a transcript. The available weight formats may also give teams options to assess deployment on the hardware environments NVIDIA lists, rather than making deployment choice solely through a hosted API.
The practical value depends on more than the feature list. Teams should test whether their languages, accents, recording conditions and specialist terms are represented well enough, and whether the output quality meets their review requirements. The one-hour recording ceiling and lack of livestream support rule out some use cases as described. The listed checkpoint formats and single-chip claim suggest several deployment options, but the source does not provide performance, memory, or cost measurements for those environments. Open weights may offer teams control over deployment, but the license and their own operating requirements still need review. Earlier coverage: NVIDIA Nemotron 3.5 Transcribe and Granary v2: Intelligent Multilingual Transcription You Can Prompt and Adapt.
Limits and uncertainty
The supplied article and report do not provide benchmark scores, evaluation methodology, error rates by language, or comparisons that would independently substantiate NVIDIA's claim of beating rivals. They also do not establish the quality of diarization, translation or timestamps in noisy audio, overlapping speech, or domain-specific recordings, nor whether every language performs equally well.
The report identifies the OpenMDW-1.1 license and checkpoint formats, but the material does not explain the license's implications for a particular commercial deployment. The source says the model handles recordings up to an hour and not livestreams; it does not detail streaming plans, batch throughput, operational latency, or the practical hardware requirements behind the single-chip claim. Granary v2's expanded language count and public materials do not by themselves demonstrate data quality, coverage balance, or fitness for a particular product domain.
Skill to build now
Create a small recorded-audio evaluation set from the languages, speaker mixes and domain vocabulary your product actually encounters. Score transcription accuracy, speaker attribution, timestamp usefulness and translation output separately, including difficult recordings with overlap or noise. Compare the result against your present pipeline before treating a unified model as a simplification. Also verify that recordings fit the stated one-hour boundary and that no live-audio requirement has been silently folded into the evaluation.
Reader takeaway
Nemotron 3.5 Transcribe is an open-weight, multilingual recorded-audio system that combines several speech tasks, and Granary v2 broadens the accompanying language dataset. That makes it worth evaluating for batch audio applications, especially where domain adaptation matters. Do not infer production quality from the feature list or NVIDIA's benchmark claim: test the languages and acoustic conditions your users actually bring, and account for the explicit lack of livestream support.