Nemotron 3.5 Transcribe at a glance
One promptable open-weight model handles transcription in 40 languages, speaker-attributed transcripts, speech translation, and questions about recordings up to an hour long, under a license that allows commercial use.
What changed with Nemotron 3.5 Transcribe
On October 28, 2026, NVIDIA introduced Nemotron 3.5 Transcribe, an open-weight speech model released on Hugging Face as NVIDIA-Nemotron-3.5-Transcribe-30B-A4B. Instead of chaining separate models for transcription, speaker labels, and translation, developers give one model plain-language instructions, such as labeling each speaker or listing the decisions made in a meeting. It transcribes 40 languages, including seven Indic languages.
The model card lists recordings of up to one hour, word timestamps, speaker-attributed transcripts, translation to and from English, and audio question answering. It accepts WAV, FLAC, MP3, and OGG files and runs with vLLM, Hugging Face Transformers, SGLang-Omni, and GGML on recent NVIDIA GPUs. The OpenMDW-1.1 license allows commercial and non-commercial use. NVIDIA says quantized versions for RTX Spark, DGX Spark, and Jetson devices are coming soon.
NVIDIA reported at launch that the model was No. 1 on the Hugging Face Open ASR leaderboard and No. 2 among open-weight models on Artificial Analysis, with about 4.3 billion active parameters. Its model card lists a 4.71% mean word error rate on English Open ASR and 15.80% on multi-speaker tests. These are vendor-reported figures.
Nemotron 3.5 Transcribe in numbers
Values are copied from the linked source on the date shown. They are not IA Compa measurements.
- English Open ASR mean word error rate (vendor-reported)as of · Source for English Open ASR mean word error rate (vendor-reported): huggingface.co (opens in a new tab)
- 4.71% WER
Where to compare Nemotron 3.5 Transcribe
Independent leaderboards and official resources for comparing this model yourself. Rankings change often; each site shows its own current data.
General leaderboards
Why Nemotron 3.5 Transcribe matters for businesses
For agencies running call recording, meeting notes, or bilingual support lines, this offers an open option that can run on your own GPU or a private cloud, so audio does not have to leave your environment. Test it on real client audio with accents, cross-talk, and industry terms before replacing a paid transcription API, and confirm your recording consent process in each state where you operate.
Put it to work: AI Voice Agents
Sources for this story
- NVIDIA RTX Spark Windows Laptops Gain New Models, Agent Tools and Games (opens in a new tab)Official
NVIDIA Blog · blogs.nvidia.com
- NVIDIA-Nemotron-3.5-Transcribe-30B-A4B model card (opens in a new tab)Official
NVIDIA on Hugging Face · huggingface.co
- Open ASR Leaderboard (opens in a new tab)Benchmark
Hugging Face · huggingface.co
Last verified: . IA Compa summarizes public announcements for business readers. We are not affiliated with the companies mentioned. Always confirm details, pricing, and availability with the official source.
- #nvidia
- #nemotron
- #speech-to-text
- #transcription
- #open-weights
- #diarization
