NVIDIA Nemotron 3.5 Transcribe and Granary v2: Intelligent Multilingual Transcription You Can Prompt and Adapt

Nemotron 3.5 Transcribe is an open-weight model that handles transcription in 40 languages, translation for 31 of them to or from English, and speaker labeling in 13, for recordings up to an hour with up to eight speakers. It takes plain-language instructions, so the same meeting file can yield a transcript, a translation, or an answer about what was decided.

NVIDIA also released Granary v2, a public manifest covering about 463,000 hours across 48 languages, plus the curation pipeline and fine-tuning recipes. In a medical case study, fine-tuning cut word error rate on held-out speakers by 45.8% relative to the best prompt.

We've seen plenty of speech models top a leaderboard. What sets this release apart is that it comes with the tools to make it fit your own audio. Prompting gets you part of the way; NVIDIA reports that simply naming the language and script cut word error rate in its medical study from 30.8% to 23.8%. Fine-tuning took it to 11.8% on the full set, and the cleaner held-out test still showed a big drop. The whole adaptation took about 14 GPU-hours on one GB300. That is within reach for a hospital system or a call center, not just a lab.

The efficiency story is strong too. Only about 4 billion of its 30 billion parameters are active per token, and it ties for first among open-weight models on Artificial Analysis's index against a model with roughly 24 billion active. Cheaper to run, open weights, and on-prem deployment options mean sensitive recordings don't have to leave the building.

Now the eyes-open part. Most leaderboard claims here are from NVIDIA's own blog, though several of the benchmarks are run by others. The medical study used a lot of synthetic speech, and NVIDIA itself notes a specialist model still beats it on exact brand-name drugs. In medicine, a misheard drug name isn't a typo; it can be a harm. NVIDIA's own advice to spot-check names, speaker changes, and answers is the right posture, and it should be the rule anywhere a transcript becomes a record.

There's a labor angle too. Strong, cheap, open transcription with speaker labels pushes more work off human transcriptionists and note-takers. That's real productivity, and it's real displacement. And easy, accurate who-said-what at scale makes recording and analyzing conversations far cheaper, which raises consent questions in workplaces and classrooms.

Our read: this is one of the more useful open releases of the season, because it's built to be adapted, not just admired. I'm watching for quantized local versions and independent checks in clinical settings.

Transcribe pairs a speech encoder with a hybrid Mamba-Transformer mixture-of-experts language model. It is built for recorded audio, not live streaming. NVIDIA says quantized versions are coming for local hardware, and the checkpoint is available on Hugging Face and ModelScope, with vLLM and SGLang-Omni support and hosted endpoints from partners.