NVIDIA has released Nemotron 3.5 Transcribe, an audio‑input language model with open weights designed for transcription. A single model handles 40 languages and recordings up to 1 hour long. A text instruction determines the requested output style for each request, and for developers it can replace multiple separate tools for speech recognition, speaker identification and alignment.
The model, which NVIDIA presents as a transcription specialist rather than a general‑purpose assistant, has 34B parameters in total, of which about 4,3B are active at each computation. Its context window reaches 65 536 tokens, i.e., the text units it can process in one go. Output options include word‑by‑word transcription, written form, attribution of speech to each speaker, inclusion of a contextual term list, and translation.
A three‑block architecture
The system combines a trainable audio encoder of 0,6B parameters that runs in parallel with a frozen speaker‑identification model, Nemotron 3 Diarization. The latter tracks up to 8 speakers, and its activity is added to the audio features every 80 ms. The model’s core is Nemotron 3.5 Lightning 30B‑A3B, a hybrid Mamba‑Transformer MoE (Mixture‑of‑Experts) architecture that activates only a subset of parameters at each step. Finally, a 50M‑parameter CTC adapter aligns the final text to the audio to produce word‑by‑word timestamps.
The model was trained through a supervised fine‑tuning phase on 256 H100 GPUs for 28 500 steps, followed by a short reinforcement‑learning phase. According to NVIDIA’s technical report, it has seen roughly 1,72 million unique audio hours, or 6,5 million hours when repetitions are counted.
Results announced by NVIDIA
All figures below are self‑reported by NVIDIA in its technical report. On the nine public English datasets of the Open ASR Leaderboard, the model records an average word error rate (WER) of 4,71 %, versus 4,94 % for Qwen3‑ASR‑1.7B, the next open model. Seven commercial systems remain ahead, including OtterAI speech‑v3.1 at 3,88 %. In multilingual evaluation across seven languages, the score is 4,62 % compared with 6,49 % for Qwen3‑ASR‑1.7B, while ElevenLabs Scribe v2 reaches 3,56 %. For multi‑speaker meetings, NVIDIA reports a rate of 15,80 % (cpWER), versus 16,83 % for DiCoW v3.3 paired with DiariZen. The company claims this is the first SALM‑type model to surpass cascade systems on this metric. A named‑entity list further reduces the error rate on those names from 12,21 % to 2,20 % on ContextASR‑Bench. Results are weaker for Mandarin in meetings: 34,95 % on AliMeeting with a distant microphone, and 32,91 % on AISHELL‑4.
Deployment and limitations
The weights are available in BF16, FP8 and NVFP4, as well as in GGUF format. The BF16 version fits on a single 80 GB H100, the FP8 model weighs roughly 35 GB and the NVFP4 about 21 GB, the latter running on an RTX 5090. The OpenMDW‑1.1 license permits commercial use, according to the source. NVIDIA also claims a speed of 2 238× real‑time on a single RTX PRO 6000 Blackwell with its NIM solution, but these optimizations are not yet open‑source, and the vLLM path available at release would be about twice as slow. The model runs only offline and is not a continuous‑stream recognition service. NVIDIA also notes weaknesses in Mandarin meetings and in language‑switching within a single sentence.
