WRSWorld Research StateContribute
WORLD RESEARCH STATE

NVIDIA: Nemotron 3 NVFP4 hybrid expert-training system

Contribution grade

Low-precision hybrid language-model training systems · Human-designed research method or resource

Reference score80,000Provisional editorial estimate of this exact contribution; its grade is assessed separately.

What changed

Implemented low-precision NVFP4 pretraining with latent-expert Mamba/attention hybrids and native multi-token prediction, then consolidated specialist agentic teachers in the Ultra post-training system.

Scientific scope

Implemented low-precision NVFP4 pretraining with latent-expert Mamba/attention hybrids and native multi-token prediction, then consolidated specialist agentic teachers in the Ultra post-training system. Super and Ultra are bundled as one realized system-family delta beyond Nano, not separate achievements by size. LatentMoE, Mamba and speculative decoding have their own prior sources; architecture and throughput are configuration-specific. Earlier family plans are not realized 2025 capabilities.

Evidence

The linked original sources describe the stated contribution and its evidence limits. WRS has not independently rerun the experiments or certified a proof. Super and Ultra are bundled as one realized system-family delta beyond Nano, not separate achievements by size. LatentMoE, Mamba and speculative decoding have their own prior sources; architecture and throughput are configuration-specific. Earlier family plans are not realized 2025 capabilities.

Contributors

Provisional WRS assessment, edition 2026-10-16.1. Independent expert calibration is pending. This grade does not certify correctness or assign ownership.

Full scientific record · Browse contribution grades · Request a correction