Nemotron 3 Nano Omni shipped open-weight on April 28. Thirty billion parameters, three billion active per token, hybrid Mamba-Transformer routing through experts on twenty-five gigs of RAM. Six multimodal leaderboards taken on day of release. On MediaPerf tagging it clears nine point nine hours of video per hour, five times GPT 5.1 throughput and six times Gemini 3.0 Pro.

The dense-scaling consensus spent two years arguing sparsity, mixed modality, and edge deployability would each cost capability. This release ships all three and beats frontier closed models on throughput while running on a workstation. Capability per dollar on agent workloads compounded another decade in twelve months.