NVIDIA Nemotron 3.5 Transcribe

Overview

Nemotron 3.5 Transcribe is NVIDIA's open-weight speech model, released on October 28, 2026. One checkpoint covers transcription, translation, speaker diarization and audio question answering. Weights are on Hugging Face and ModelScope as NVIDIA-Nemotron-3.5-Transcribe-30B-A4B.

The architecture is a hybrid Mamba-Transformer mixture-of-experts language model with a speech encoder, 30 billion total parameters and about 4 billion active per token. Language coverage differs by task: 40 languages for transcription, 31 for translation to and from English, 13 for diarization, and English only for audio question answering. A single recording can run up to one hour, carry up to eight labelled speakers, and produce word-level timestamps.

Artificial Analysis reports 2.8 percent AA-WER, tied for first among open-weight models, while the model activates about 4 billion parameters against 24 billion for the comparison model. On the Hugging Face Open ASR Leaderboard it leads open weights on all three tracks: 5.33 percent average WER in English, 4.62 percent multilingual and 10.98 percent on long audio.

Key Features

• Four tasks in one checkpoint: Transcription, translation with English across 31 languages, diarization across 13 languages and English audio QA all run from the same weights. The prompt picks the job, for example asking for speaker labels or asking directly what decisions a meeting reached.
• 2.8 percent WER at 4B active: AA-WER of 2.8 percent ties for first among open weights. On the AA-AgentTalk set it reaches 1.8 percent against 2.3 percent for the next open model. Parameter efficiency matters more here than the absolute score.
• Long audio and multi-speaker meetings: Up to one hour per recording and eight speakers. The long-audio track lands at 10.98 percent average WER, the lowest in its comparison set. Diarization reaches 18.26 percent DER on DAI-ASR-I18N, lowest of 14 systems.
• 100 hours of audio in three minutes: On a single RTX PRO 6000 Blackwell Server Edition using the upcoming NVFP4 quantisation, the model processes 100 hours of audio in three minutes when fed 30 second clips at high concurrency.
• Medical fine-tuning with published numbers: On the Eka Care medical benchmark the out-of-box WER is 30.8 percent, 23.8 percent with the best prompt, and 11.8 percent after fine-tuning, with 12.17 percent on held-out speakers. Fine-tuning took about 14 GPU hours on one GB300.
• Granary v2 alongside: The matching dataset ships at 463,062 hours across 48 languages from 16 sources, built with NeMo Curator. Each record carries three transcript variants: verbatim, inverse-normalised with disfluencies, and inverse-normalised without them.

Use Cases

• Meeting and interview transcription, one hour per file with up to eight speakers and labels.
• Cross-language archiving, translating 31 languages into English before indexing.
• Contact-centre quality review using word-level timestamps to line up script hits.
• Clinical notes, where the published fine-tune path reaches 11.8 percent WER and 12.17 percent on held-out speakers.
• Audio search, asking a recording what decisions were made instead of reading a transcript.
• Teams that need data to stay inside their own network, self-hosting through vLLM or SGLang-Omni.

Pros

• 2.8 percent AA-WER, tied first among open weights, at about 4B active parameters.
• One checkpoint covers transcription, translation, diarization and audio QA.
• Handles one hour of audio and eight speakers, with the lowest long-audio WER in its set.
• 100 hours of audio in three minutes on one GPU with NVFP4.
• Fine-tuning recipe and cost are public, with 12.17 percent held-out WER on the medical set.
• Weights on Hugging Face and ModelScope, with vLLM and SGLang-Omni supported at launch.

Pricing

Open weights, free to download and run, with commercial use governed by NVIDIA's open model licence. Self-hosting costs only hardware, and vLLM and SGLang-Omni support the model at launch with a NeMo-Speech.cpp local build following. Baseten, DeepInfra and Simplismart host managed endpoints at their own rates. Granary v2 is public as well, though source audio must be fetched under its original licences.

Summary

The value here is collapsing a pipeline into one call. Meeting transcription used to mean chaining an ASR model, a diarizer and a translator, then aligning timelines across all three. A single checkpoint with a prompt covers it. Thirty billion parameters with four billion active keeps it on one GPU and the fine-tune budget at around 14 hours. The design target is recorded audio; real-time streaming is a different problem.

Version History

• Nemotron 3.5 Transcribe open-sourced: transcription, translation and diarization in one model (2026-10-28): NVIDIA released open weights for NVIDIA-Nemotron-3.5-Transcribe-30B-A4B, a hybrid Mamba-Transformer mixture of experts with a speech encoder at 30 billion total and about 4 billion active parameters. It transcribes 40 languages, translates 31 to and from English, diarizes 13 languages and answers English audio questions, handling one hour per recording and up to eight speakers with word-level timestamps. Artificial Analysis AA-WER is 2.8 percent, tied first among open weights, and Hugging Face Open ASR shows 5.33 percent English, 4.62 percent multilingual and 10.98 percent long audio. Granary v2 ships alongside at 463,062 hours across 48 languages. vLLM and SGLang-Omni support it at launch.