November 1, 2026 | AI | By Mahesh
Nvidia released Nemotron 3.5 Transcribe on October 28, 2026, an open-weight speech model that transcribes 40 languages, labels up to eight speakers, translates 31 languages to or from English and answers questions about recordings up to one hour long [1][2]. Nvidia reports an average word error rate of 5.33% on the English track of Hugging Face's Open ASR Leaderboard, the top result among open models [1][2]. The weights are free and the license permits commercial use, but the model handles recorded audio only [1].
- Nemotron 3.5 Transcribe is an open-weight model that transcribes 40 languages, labels up to eight speakers and translates 31 languages to or from English.
- It averages a 5.33% word error rate on the English Open ASR Leaderboard, and it is a mixture-of-experts model with about 4 billion active parameters.
- Weights are free and permit commercial use, but it handles recorded audio only and the quantized single-GPU builds are not out yet.
What Nemotron 3.5 Transcribe is and what is new
One checkpoint handles several jobs that usually need separate tools: transcription, speaker labels, translation and spoken-question answering [2]. It accepts plain-language prompts, such as a request for a clean or verbatim transcript or a list of medical or product terms to spell correctly [1]. It attributes speakers in 13 languages and can separate up to eight voices [2].
Nvidia also released Granary v2, a speech dataset of 463,062 hours across 48 languages, up from 25 in the first version [1]. The dataset is free, but users fetch the audio from each original source under that source's license, so it is not a single download [1].
Under the hood: size and architecture
SQ Magazine describes a 30-billion-parameter hybrid Mamba-Transformer mixture-of-experts model with about 4 billion parameters active per token [2]. The Next Web gives 34 billion total parameters and notes that the Hugging Face model name reads 30B-A4B, so the total differs between sources and the model card should settle it [1]. The Next Web adds that the language component builds on Nemotron 3.5 Lightning, released in August [1].
Developers can serve the current checkpoint with vLLM today, and local deployment through NeMo-Speech.cpp is pending [2]. Nvidia says the model can run on a single GPU, but the quantized builds that requires are marked coming soon [2]. For another recent open release, see Mistral Large 4 Preview: 1.05T Open-Weight MoE, Specs and Pricing.
Benchmarks: leaderboard and vendor figures
The 5.33% average word error rate comes from the English track of the Open ASR Leaderboard, where Nvidia says Nemotron 3.5 Transcribe ranks first among open models [1][2]. SQ Magazine adds that Artificial Analysis puts it at 2.8% on its AA-WER index, tied for first among open-weight models [2]. For speaker labeling, the model posted an 18.26% diarization error rate on David AI's multilingual track, narrowly ahead of Nvidia's own Nemotron 3 Diarization model at 18.38% [2].
| Measure | Result | Context |
|---|---|---|
| Open ASR Leaderboard, English average WER | 5.33% | Top among open models, per Nvidia |
| AA-WER index (Artificial Analysis) | 2.8% | Tied first among open-weight models |
| Diarization error rate, David AI multilingual | 18.26% | Nemotron 3 Diarization at 18.38% |
| Medical fine-tune WER, held-out speakers | 12.17% | Down from 22.44% before tuning |
The fine-tuning case study used medical speech from Indian doctors and cut errors by 45.8% on held-out clips in about 14 hours on one GPU [1][2]. SQ Magazine notes that general English accuracy slightly declined after tuning [2]. Nvidia's speed claim of 100 hours of audio in under three minutes depends on the unreleased quantized builds [2].
Pricing, availability and how to access it
The weights are free under an open-weight license that permits commercial use, and the model page is on Hugging Face as Nvidia Nemotron 3.5 Transcribe 30B-A4B [1]. The Next Web compares this with a hosted alternative, SpaceXAI's Grok transcription at 10 cents per hour [1]. Self-hosting removes that fee, but teams then pay for their own GPUs.
Limitations, trade-offs and alternatives
The model is built for recorded audio and not live transcription [1][2]. Translation covers 31 languages against 40 for transcription, and the headline accuracy figure is English only, so results in other languages may differ [1]. Most figures come from Nvidia or leaderboards Nvidia cites, and I found no independent test of the multilingual claims.
Teams that need live captions should look elsewhere for now. Teams with recorded calls, interviews or lectures can test the checkpoint on their own audio, and those with specialist vocabulary can try the fine-tuning recipes that ship with it [2]. For another recent model release with disclosed specs, see Claude Haiku 5.5 Launch: $0.10 Input, 1M Context, Benchmarks.
Sources
- The Next Web. "Nvidia releases an open speech model that transcribes 40 languages." thenextweb.com
- SQ Magazine. "Nvidia open Nemotron 3.5 Transcribe model." sqmagazine.co.uk
Read More on Depth Grid
- Nvidia in Talks to Deepen Stake in or Buy Reflection AI
- Claude Haiku 5.5 Launch: $0.10 Input, 1M Context, Benchmarks
- Meta Muse Wonder Launch: 1M Context, $4 API Price, Benchmarks
- Mistral Large 4 Preview: 1.05T Open-Weight MoE, Specs and Pricing
Article by Mahesh | Depth Grid
Discussion (0)
Comments are disabled for this post.