NVIDIA releases Nemotron 3.5 Transcribe, an open model that transcribes 40 languages and labels speakers

NVIDIA has released Nemotron 3.5 Transcribe, an open-weight speech model that turns recorded audio into text. The same checkpoint transcribes, translates, marks who is speaking, and answers questions about a recording. This suggests that an app chaining separate models for those jobs could try a single one instead. The model went up on Hugging Face on October 28, 2026, under the OpenMDW-1.1 license, which permits commercial use.

What one model does

The model card lists transcription in 40 languages and translation between English and 31 of them, in both directions. In 13 languages the model also performs speaker attribution, known as diarization: it marks which person said which words. NVIDIA says it handles recordings up to an hour long and labels up to eight speakers.

Output is steered with a plain-language instruction rather than with settings. One recording can come back as a transcript with speaker labels, as an English translation, or as an answer to a question such as which decisions a meeting reached. A prompt can also carry a list of expected terms, such as drug names, which the model is told to use only when the audio supports them. Question answering was trained and evaluated only in English, and the card warns it may not work reliably in other languages.

This is an offline model. NVIDIA built it for recorded audio, not for live streaming, so real-time captioning is not what it is meant for.

What it takes to run

The model has 34 billion parameters in total, and about 4.3 billion of them, including the audio encoder, are active for each token. That is the mixture-of-experts design: each step runs only part of the network, which cuts the amount of computation, but all of the weights still have to fit in memory.

The reference weights are in BF16, and NVIDIA's quick start serves them on one H100 with 80 GB of memory. The card also names FP8 and NVFP4 builds, quantized versions that store the weights at lower precision so they need less memory, and its hardware table pairs them with the RTX 5090 and DGX Spark. NVIDIA's announcement says those versions are coming soon, and on October 28 their repositories were not public. Until they ship, running the model yourself is a job for a data-center GPU, not a desktop.

What the measurements show

NVIDIA reports the lowest average word error rate among open-weight models on the multilingual Open ASR Leaderboard, 4.62% across seven languages. Word error rate is the share of words a transcript gets wrong, so lower is better. On the speech-to-text ranking from Artificial Analysis, an independent benchmarking firm, NVIDIA says the model ties for first among open-weight models at 2.8%, with roughly a sixth of the active parameters of the model it ties.

NVIDIA also published a fine-tuning case study on medical speech from Indian doctors and medical students. On speakers held out from tuning, the error rate fell from 22.44% with the best prompt to 12.17% after fine-tuning, a 45.8% relative reduction, and the whole adaptation took about 14 GPU-hours. The same tuning raised the error rate on general English slightly, from 1.54% to 1.69%.

Where Serbian stands

Serbian is not among the 40 languages. Croatian is, rated production-ready by NVIDIA, with a word error rate of 8.31% on the FLEURS test set, and Slovenian and Bulgarian are listed as well. NVIDIA has published no results for Serbian, so how much the Croatian support helps with Serbian speech is unknown.