Nvidia releases Nemotron 3.5 Transcribe, a free 30 billion parameter speech model

Nvidia just gave away, for free, a transcription model good enough to worry companies that charge for one. That's the real story behind Wednesday's release.

Nvidia released Nemotron 3.5 Transcribe on October 28. It's a 34 billion parameter open-weight speech model, including its audio encoder, that transcribes in up to 40 languages, translates 31 of them to and from English, and tells up to eight speakers apart in a single recording. According to TechCrunch, the model is built to run on a single chip, from a Jetson AGX Thor edge device up through Nvidia's DGX Spark and Blackwell datacenter GPUs. It ships under the OpenMDW-1.1 license, with checkpoints available in BF16, FP8, NVFP4, GGUF and DFlash2 formats. So developers can run it at whatever precision their hardware allows.

The architecture is more interesting than the headline parameter count suggests. Nemotron 3.5 Transcribe combines a Transformer speech encoder with a frozen Nemotron 3 Diarization model and the Nemotron 3.5 Lightning 30B-A3B model. That last one is a hybrid Mamba-Transformer mixture-of-experts system that only activates around 4.3 billion parameters at inference time. It's a mixture-of-experts trick borrowed straight from Nvidia's large language model work, repurposed for audio. The company says the design makes it the first speech-audio language model, or SALM, to beat specialized multi-stage pipelines at speaker-attributed transcription in overlapping, multi-speaker audio: a notoriously hard problem for any ASR system.

On Hugging Face's Open ASR Leaderboard, Nvidia reports a 4.71% mean word error rate across nine English test sets and 4.62% on the multilingual leaderboard. The company says those numbers put it ahead of open-weight rivals. A lightweight CTC adapter aligns the generated transcript back to the audio after decoding. That's what lets the model hand back word-level timestamps tied to specific speakers, even across an hour-long recording. On an RTX PRO 6000 Blackwell GPU, Nvidia claims throughput of up to 2,238 seconds of audio processed per second of compute.

None of this sells a chip by itself. What it does is make Nvidia's hardware the obvious place to run the best open transcription model available. That's exactly the playbook Nvidia has run before with Parakeet and Canary, and with its Granary speech datasets, whose first version included around a million hours of multilingual audio when it debuted in 2025. Give away the software, make your silicon the natural home for it. Enterprises that build meeting-transcription or medical-dictation tools on Nemotron 3.5 Transcribe can run it on Nvidia hardware instead of paying a third-party transcription API.

That matters more right now because the voice AI market is unusually crowded. TechCrunch notes the release landed in the same stretch as Mistral's Voxtral 3, ElevenLabs' Realtime product for instant speech, and Salesforce's Launa-Omni, all shipping within weeks of each other. Nvidia's pitch is narrower than it sounds: Nemotron 3.5 Transcribe works on recorded audio, not live streams. That puts it squarely against business-meeting and medical-transcription use cases, rather than the real-time voice-agent market ElevenLabs is chasing with Realtime. Developers can also fine-tune it on industry-specific vocabulary using Nvidia's NeMo Speech tools, aimed at exactly the kind of jargon-heavy dictation that generic models fumble.

Here's the part that should worry smaller speech-AI vendors. OpenAI's Whisper, released back in 2022, already turned transcription from a differentiated product into a commodity API call at $0.006 a minute. A free, state-of-the-art, speaker-aware model with translation built in pushes that commodification further down the stack. The per-minute API fee disappears for anyone running it on hardware they already operate, although compute and operating costs remain. Deepgram sits at an estimated $100 million revenue run rate as of this past July, and ElevenLabs reportedly hit $600 million in June. Both have leaned on real-time voice infrastructure and broader audio platforms rather than raw transcription accuracy to stay ahead of exactly this kind of commoditization. Nemotron 3.5 Transcribe doesn't touch their real-time voice-agent business. It does take direct aim at the batch-transcription tier underneath it: the part of the market with the thinnest margin and the least defensible moat.

Nvidia isn't trying to become a speech-AI company. It's trying to make sure nobody can build a speech-AI business on top of anyone else's chips without first checking what Nvidia is giving away for free.

Also read: FuriosaAI's valuation jumps fourfold to $2.9 billion as it chases Nvidia • Samsung joins SK Hynix and Micron declaring AI memory sold out for years • Nscale is seeking a $30 billion IPO as investors push back hard

This article is posted in AI News, check it out for more related stories.