Nvidia Nemotron 3.5 Transcribe: open speech-to-text, no streaming

Nvidia released Nemotron 3.5 Transcribe on October 28: a free, open-weight speech-to-text model that handles 40 languages and recordings up to an hour long, with a license that allows commercial use. If you pay per minute to transcribe sales calls, support calls or meetings, you now have a self-hostable option from a vendor that isn't going away. One catch up front: it doesn't stream. It's built for recordings, not live calls.

What actually happened

TechCrunch reported the launch, and the details are in Nvidia's Hugging Face model card:

• Size: about 34B total parameters, roughly 4.3B active per token (a mixture-of-experts design, which is why the name says "30B-A4B").
• License: OpenMDW-1.1. The card states it's ready for commercial or non-commercial use.
• Languages: 40 for transcription. Audio reasoning and spoken-instruction features were evaluated in English only.
• Length: recordings up to one hour, with speaker tags in the output. The training data includes mixes of up to eight speakers.
• Input: WAV, FLAC, MP3, OGG, or anything that decodes to 16 kHz mono PCM.
• Streaming: none. The card calls it an offline model.
• Hardware: the quick start targets a single H100 80GB. FP8 and NVFP4 quantized variants are listed for smaller footprints.
• Nvidia's numbers: 4.71% mean word error rate on the English Open ASR Leaderboard, 4.62% on the multilingual set, and 15.80% on hard multi-speaker sets like meeting recordings. These are vendor-reported.

Nvidia also shipped an updated open speech dataset, Granary v2, and its NeMo Speech tools so teams can tune the model on their own industry terms.

Why it matters for your business

Most small businesses already record calls. Few do anything with those recordings, because transcription at volume costs real money and every vendor wants a per-minute fee. An open model with a commercial license changes the math. You run it once a night against yesterday's calls, then feed the text to whatever does the summarizing, tagging and CRM updates.

The batch-only design is the honest limit. If you want a voice agent that answers the phone, this is the wrong tool; you need a streaming model for that. For call-log analysis, meeting notes and compliance archives, batch is fine, and usually cheaper.

Do the math before you switch. A rented H100 isn't free, and hosted transcription APIs now charge cents per hour of audio, so at low volume the API can still win. Price your own monthly hours against GPU rental before you move anything. Test it on 50 of your own recordings, with your own accents and product names, before you trust the leaderboard.

Sources: Hugging Face: NVIDIA Nemotron 3.5 Transcribe model card, TechCrunch.