Nemotron 3.5 Transcribe
nvidia · Released Oct 2026
Conditional. Nemotron 3.5 Transcribe is NVIDIA's open-weight offline speech model under OpenMDW-1.1, transcribing 40 languages, translating 31 to or from English and attributing speakers in 13, for recordings up to an hour. NVIDIA says it ties for first among open-weight models on Artificial Analysis at 2.8% AA-WER, but that is NVIDIA's own summary, it is not for live streaming, and no hosted price was read. Suits teams self-hosting transcription of meetings and calls.
Is it right for you?
Good for
- Self-hosted transcription of long recordings; the model card says it is trained on and handles recordings up to one hour, with a 65,536-token context
- Multi-speaker audio; NVIDIA says it labels up to eight speakers and attributes speakers in 13 languages (NVIDIA's claim, untested by us)
- Multilingual work with English translation; NVIDIA says it transcribes 40 languages and translates 31 of them to or from English
- Efficient serving of an open model; the blog says about 4B active parameters, and the card counts 4.3B active of 34B including the audio encoder
Not good for
- Live streaming transcription; the blog says it is built for recorded audio, not live streaming, and the card calls it an offline model
- Comparing hosted prices; the blog names Baseten, DeepInfra and Simplismart as endpoints, and the pages we read state no prices
- Decisions that need independent accuracy confirmation; leaderboard figures are NVIDIA's summary of third-party boards, and the blog's 5.33% and the card's 4.71 for the Open ASR Leaderboard differ, which we did not reconcile
How it performs by task
Transcription accuracy
NVIDIA says it ties for first among open-weight models at 2.8% AA-WER on Artificial Analysis; NVIDIA's summary of a third-party board, untested by us
Speaker diarization
attributes speakers in 13 languages and labels up to eight speakers per NVIDIA; the diarization benchmark figures were read once and are not stored, untested by us
Speech translation
translates 31 of its 40 languages to or from English per NVIDIA; no translation score read
Pricing
Input
Not stated
Output
Not stated
Context
Open weights, OpenMDW-1.1; no hosted price read
Benchmarks
| Benchmark | Score | Source |
|---|---|---|
| Artificial Analysis AA-WER, open-weight models (NVIDIA-reported) | 2.8% (tied first) | Source |
No verdict changes yet
The clock starts day one. Changes land here as our verdict evolves.
Sources
Verification log
- Status· No changes
Automated agent
- Pricing· No changes
Automated agent