NVIDIA Nemotron 3 Nano Omni: A 30B multimodal model with 3B active parameters
NVIDIA's Mamba-Transformer-MoE model accepts images up to 1,840 × 1,840, audio up to 20 minutes, and long text. It scored 65.8 on OCRBenchV2-En, 74.1 on DailyOmni, and 89.4 on VoiceBench. NVIDIA reported 9× the throughput of comparable open omni models. BF16, FP8, and NVFP4 checkpoints are on Hugging Face, and the model is also available via OpenRouter and NVIDIA.
Visit original source