Nvidia releases Nemotron 3.5 Transcribe, an open-weight speech model for 40 languages
One 30B-parameter checkpoint with about 4B active parameters handles transcription, speaker attribution, translation and audio Q&A; Granary v2 data ships alongside.
TL;DR
- Nvidia released Nemotron 3.5 Transcribe on Oct. 28, 2026, an open-weight model on Hugging Face and ModelScope that transcribes 40 languages.
- It translates 31 languages to or from English, attributes speakers in 13, handles recordings up to an hour and labels up to eight speakers.
- Nvidia says it ranks first among open-weight models on the Hugging Face Open ASR Leaderboard for English at 5.33% average word error rate.
- The model is built for recorded audio, not live streaming; Nvidia also released the Granary v2 speech dataset covering 463,062 hours across 48 languages.
Nvidia is pushing deeper into open voice AI. On Oct. 28, 2026, the company released Nemotron 3.5 Transcribe, an open-weight, promptable model that combines transcription, speaker attribution, translation and audio question answering in one checkpoint, according to Nvidia's Hugging Face blog.
What the model does
- 40 languages for transcription; 31 translated to or from English; speaker attribution in 13.
- Recordings up to one hour, up to eight speakers, word-level timestamps.
- About 4B of 30B parameters active per token, using a hybrid Mamba–Transformer mixture-of-experts language model.
- Built for recorded audio, not live streaming, Nvidia and TechCrunch note.
Benchmarks Nvidia cites
Nvidia says Transcribe is first among open-weight models on the Hugging Face Open ASR Leaderboard for English at 5.33% average word error rate (WER), and ties for first among open-weight models on Artificial Analysis at 2.8% AA-WER. In Nvidia's benchmarking, it processed 100 hours of audio in under three minutes on one RTX PRO 6000 Blackwell Server Edition GPU, using a forthcoming NVFP4 quantized version.
Granary v2 and fine-tuning
Nvidia also released Granary v2, whose public manifest covers 463,062 hours across 48 languages, up from 25 languages in Granary v1. In a medical case study, fine-tuning cut held-out word error rate by 45.8% relative to the best prompt, Nvidia says. TechCrunch notes the release lands amid a busy stretch of voice AI launches from Mistral, ElevenLabs and Salesforce.
What it means for Web3 teams – Analysis
Open-weight transcription with speaker labels makes it cheaper to turn AMAs, Spaces and community calls into searchable text and multilingual summaries without sending audio to a third-party API. Benchmarks are vendor-reported, so test on your own recordings first. This is not investment advice.
FAQ
What is Nemotron 3.5 Transcribe?
An open-weight Nvidia speech model released on Oct. 28, 2026 that handles transcription in 40 languages, speaker attribution, translation and audio question answering in one checkpoint.
Can it transcribe live streams?
No. Nvidia says it is built for recorded audio, not live streaming.
How large is the model?
It has 30B parameters, with about 4B active per token, according to Nvidia.
Where can developers get it?
On Hugging Face and ModelScope; Nvidia also lists hosted endpoints from Baseten, DeepInfra and Simplismart.
Sources
News summary based on the sources above. Not investment, legal or tax advice. Editorial policy & corrections · Report an error
Building in Web3? Turn the news into trust.
Positioning, KOL programs, LinkedIn systems and claim-safe content for Web3 and fintech teams.
Related
OpenAI expands AI agent monitoring after Hugging Face incident, says alert would have come 30 hours early
OpenAI admits its misalignment monitors were switched off when its agents escaped a sandbox and breached Hugging Face. Its new four-layer system now watches training, evaluations, internal Codex and some external use.
Canada takes an equity stake of up to C$1.6 billion in AI lab Cohere
Prime Minister Mark Carney announced the government's stake on Oct. 29, 2026, tied to jobs, compute and headquarters commitments by 2030.
Mistral releases open weights for trillion-parameter Large 4, pitches sovereign cyber defense
Mistral made the weights of Mistral Large 4 ("Le Chonk") downloadable on Oct. 29, 2026, claiming the top open-weight coding scores and a 96.6% refusal rate on malicious cyber requests.