Nvidia Nemotron 3.5 Transcribe: open model for 40 languages

Nvidia has released Nemotron 3.5 Transcribe, a free open-weight speech recognition model, together with Granary v2, a large open speech dataset and the pipeline used to build it. The model is available on Hugging Face and ModelScope, and Nvidia is pitching it at developers who build meeting, call-centre and medical transcription tools.

What the model does

Instead of separate tools for each job, Nvidia puts several speech tasks into one checkpoint that you steer with a plain-language prompt, for example "transcribe this recording and label each speaker" or "translate this audio into English". According to Nvidia:

• it transcribes 40 languages and translates 31 of them to or from English;
• it can tell up to eight speakers apart (speaker attribution works in 13 languages);
• it accepts recordings up to one hour long and supports word-level timestamps;
• it can answer questions about a recording (audio Q&A is trained in English).

One clear limit: it is built for recorded audio, not live streaming, so it isn't a drop-in engine for real-time captions or voice assistants.

How big is it, and where does it run?

The model has 30 billion parameters, but it uses a hybrid Mamba–Transformer mixture-of-experts design, so only about 4 billion are active per token. Nvidia says it runs on a single GPU, from the Jetson AGX Thor robotics module and the DGX Spark desktop box up to Blackwell data-centre cards. It can be served with vLLM or run locally with NeMo-Speech.cpp. If you're wondering what hardware open models need at home, our guide to how much VRAM you need for local AI is a good starting point.

Benchmarks (Nvidia's figures)

Nvidia says the model ranks first among open-weight models on Hugging Face's Open ASR Leaderboard for English, with a 5.33% average word error rate, and ties for first among open models on Artificial Analysis's speech-to-text index at 2.8%. These are the company's own claims based on public leaderboards; real-world accuracy will vary with accents, noise and jargon, and Nvidia itself advises spot-checking names and speaker changes.

Granary v2 and fine-tuning

Granary v2's public manifest covers about 463,000 hours of speech in 48 languages, up from 25 languages in the first version. Nvidia also published fine-tuning recipes: in a medical case study, adapting the model cut the word error rate on held-out speakers by 45.8% compared with the best prompt alone.

As TechCrunch notes, the release lands in a crowded month for voice AI. For the bigger picture on what "open weights" means and what you can legally do with such models, see our explainer on open-weight AI models.