Key takeaways
- NVIDIA released Nemotron 3.5 Transcribe on October 28 as open weights under the OpenMDW-1.1 license.
- It transcribes up to an hour of audio in 40 languages and uses about 4.3B active parameters.
- NVIDIA reports a 4.71% mean English word error rate, ahead of Qwen3-ASR-1.7B at 4.94%.
NVIDIA released Nemotron 3.5 Transcribe on October 28, an open-weight speech model that turns recordings of up to one hour into text in 40 languages. According to the model card on Hugging Face, it has the lowest error rate among open-weight models on the Open ASR Leaderboard, a public speech-recognition benchmark. Those numbers come from NVIDIA.
The weights are free to download under the OpenMDW-1.1 license, which the card says allows commercial and non-commercial use. That makes it a direct alternative to open speech models such as Whisper, Parakeet and Qwen3-ASR for teams that want to run transcription on their own hardware.
One model, several kinds of transcript
Most speech models return one kind of text. This one takes an instruction with the audio and changes its output to match. The card lists verbatim transcripts, written form with punctuation and numerals, transcripts guided by a list of names or terms, speaker-labeled transcripts, translation, language identification and questions about the audio. Word timestamps are available too.
It is a mixture-of-experts model, which runs only part of its network for each token. The total is 34 billion parameters including the audio encoder, and about 4.3 billion are active at a time. It reads up to 65,536 tokens of context and is built on NVIDIA's Nemotron 3.5 Lightning language model. NVIDIA recommends one 80 GB H100 GPU.
What NVIDIA reports
The card says results were captured on October 12. The figures are word error rates, so lower is better.
| Test | Nemotron 3.5 Transcribe | Next best open-weight model |
|---|---|---|
| English, nine Open ASR sets | 4.71% | Qwen3-ASR-1.7B, 4.94% |
| Long recordings, Open ASR | 10.98% | Parakeet-TDT-0.6B-v3, 13.37% |
| Seven languages, multilingual Open ASR | 4.62% | Qwen3-ASR-1.7B, 6.49% |
On meetings with several speakers, NVIDIA reports a 15.80 mean error rate against 16.83 for the next best published system. The gaps are small on English and larger on long recordings and other languages.
Speed and limits
NVIDIA says its optimized serving stack, the Nemotron Transcribe NIM, processes up to 2,238 seconds of audio per second on one RTX PRO 6000 Blackwell GPU, using a 4-bit version of the model with speculative decoding. Those optimizations are not in the open-source release yet. The card says the code that ships is for development and is roughly two times slower.
This is an offline model. The card says it is not a real-time streaming service, so it fits recorded meetings, calls and media, not live captions. The weights come in BF16, FP8 and NVFP4 versions.
NVIDIA also released Granary v2, a speech dataset it says covers 463,062 hours across 48 languages. Independent tests on real recordings will show whether the leaderboard lead holds up.
Sources
- NVIDIA-Nemotron-3.5-Transcribe-30B-A4B model cardNVIDIA on Hugging Face
- NVIDIA NeMo SpeechNVIDIA
- OpenMDW License 1.1OpenMDW
