Nvidia releases open Nemotron 3.5 Transcribe for 40 languages

Nvidia released Nemotron 3.5 Transcribe on 28 October 2026, a free open-weight model that turns recordings into text and can be told what to do with them. According to its Hugging Face model card, it transcribes 40 languages, takes recordings up to one hour, and can return a plain transcript, a speaker-labelled one, a translation or an answer to a question about the audio. It is released under the OpenMDW-1.1 licence, which the card describes as ready for commercial or non-commercial use.

What it does

The card describes one speech-and-language model steered by a text instruction. The same recording can become a verbatim transcript, a cleaned-up one, a speaker-attributed one, or an English translation. Users can pass in context such as product or drug names as hints. Nvidia lists word timestamps and language identification among the tasks.

The Decoder, citing Nvidia, reports translation for 31 languages to or from English, speaker identification in 13 languages, and up to eight speakers per recording. A Decoder headline says 30 billion parameters with about 4 billion active; the model card's own table says 34B total with 4.3B active, including the audio encoder. Use the card's figures.

What Nvidia says it scores

• English: Nvidia reports a mean word error rate of 4.71 on nine public sets of the Hugging Face Open ASR Leaderboard, against 4.94 for Qwen3-ASR-1.7B, the next best open-weight model it lists.
• Long recordings: 10.98 on the leaderboard's long-form set, against 13.37 for Parakeet-TDT-0.6B-v3.
• Meetings with several speakers: Nvidia reports a mean concatenated word error rate (cpWER, which also penalises wrong speaker labels) of 15.80 across four meeting datasets, against 16.83 for the next best published system.
• Artificial Analysis, an independent evaluator quoted by The Decoder, puts it at 2.81 percent error, close to Mistral's Voxtral Small at 2.77 percent with about a sixth of the active parameters.

Most of these are Nvidia's own runs on public benchmarks. The Artificial Analysis comparison is the one figure here from a party other than the vendor, and it shows the model near the top rather than ahead.

What it takes to run

Nvidia's quick-start serves the model with vLLM on one H100 80 GB GPU, and the card lists BF16, FP8 and NVFP4 versions. The card also says the speed numbers it publishes, up to 2,238 seconds of audio transcribed per second on one RTX PRO 6000 Blackwell GPU, come from Nvidia's optimised NIM container, and that these optimisations are not yet in the open-source release. The open version is expected to be roughly two times slower. Plan capacity on that, not on the headline.

Adapting it to your jargon

Alongside the model Nvidia released Granary v2, a speech dataset of about 844,000 hours in 48 languages, and tooling to fine-tune on your own recordings. The Decoder reports an Nvidia test on recordings of Indian doctors and medical students: a more precise prompt cut the error rate from 30.8 to 23.8 percent, and fine-tuning brought it to 11.8 percent, in about 14 GPU hours on one GB300. The same report says the error rate on general English speech rose slightly after specialising and that a dedicated medical model stayed better on brand-name drugs. Nvidia recommends holding back at least 100 real, verified recordings for testing.

What the sources leave out

It is offline only. TechCrunch notes live use stays with rivals such as ElevenLabs, and the card says the model is built for recorded audio. The card's limitations list substitutions, deletions, repetitions, hallucinated speech, wrong language choice and speaker attribution errors, and recommends targeted testing on long recordings, overlapping speakers, noise, names and code-switching. Audio question answering was evaluated only in English. No price applies to the weights; the cost is your GPUs. None of the sources reports results on your audio.