NVIDIA's open-weight mixture-of-experts speech model for offline transcription in 40 languages, up to one hour of audio.
Updated EU status checked Last shown
Sources: the vendor's own pages (linked above) and creator videos at the linked timestamps. EU status is information, not legal advice; see how we check.
Nemotron 3.5 Transcribe is an open-weight offline speech and audio model from NVIDIA, built on the Nemotron 3.5 Lightning 30B-A3B base. It uses a hybrid mixture-of-experts architecture with roughly 30B total parameters and about 4.3B active per token, including the audio encoder. It supports a 65,536-token context and recordings up to one hour.
A single model handles several tasks selected through text instructions: transcription in a chosen style, word timestamps, speaker diarization, language identification, code-switching, speech translation and audio question answering. The model card lists 40 languages, including most EU languages (e.g. Czech, German, French, Italian, Polish, Maltese, Irish-excluded) as well as Hindi, Japanese, Korean, Arabic and Chinese. Training data includes NVIDIA's Granary-v2 dataset.
Weights are published on Hugging Face in BF16, FP8 and NVFP4 variants, under the OpenMDW-1.1 license. NVIDIA provides NeMo Speech for fine-tuning and serving, SGLang-Omni serving instructions, a technical report and a live Hugging Face demo. It is designed for offline (non-streaming) use and needs GPU hardware sized for a ~30B-parameter model.
Watch the moment: the exact second top creators show it. Videos load only when you press play, via youtube-nocookie.com.
Speech recognition model with controllable transcript style, 40 languages, speaker identification and word timestamps, up to an hour of audio. About 34B parameters with 4.3B active.
Closest neighbours by meaning, not just by category.
Model · Nvidia
NVIDIA's family of open models for agentic AI: reasoning, vision, speech, retrieval and safety models.
Model · NVIDIA
NVIDIA's open multimodal model (120B total, 12B active parameters) for image, video and document understanding.
Model · NVIDIA
NVIDIA's open-weight ~550B mixture-of-experts LLM with a hybrid Mamba-transformer design, aimed at agentic workloads.
Model · NVIDIA
NVIDIA family of multilingual text embedding models (8B and 1B) for retrieval, RAG and coding agents.
Model · NVIDIA
NVIDIA's open-weight 30B mixture-of-experts model with 1M-token context, aimed at long-running agent workloads.
Model · Google
Google's speech-to-text model with verbatim and smart modes, speaker detection, word timestamps and a live streaming variant.
