Descriptions:
NVIDIA has released Nemotron 3 Super, a 120-billion parameter open-source AI assistant trained on 25 trillion tokens, alongside a 51-page research paper that fully documents its training data, methodology, and architecture. In this Two Minute Papers episode, Dr. Károly Zsolnai-Fehér unpacks the four technical innovations that make the model stand out: NVFP4 quantization, multi-token prediction, Mamba-based memory layers, and stochastic rounding.
The headline result is speed: the NVFP4 variant runs up to 7 times faster than comparably capable open models while maintaining equivalent accuracy. NVFP4 achieves this by selectively rounding numerical computations—leaving sensitive calculations intact and compressing only the parts where precision loss is negligible. Multi-token prediction generates 7 tokens simultaneously rather than sequentially, adding another significant throughput gain. Mamba layers replace repeated context re-reads with a compressed memory system that retains key information and discards filler, enabling efficient handling of long inputs. Stochastic rounding corrects for cumulative quantization error by introducing zero-averaged noise, ensuring accuracy holds across hundreds of generation steps.
While Nemotron 3 Super roughly matches top closed-source frontier models from about 18 months ago, its combination of full transparency, free availability, and dramatically faster inference positions it as a meaningful benchmark for the open-source AI ecosystem—and the accompanying research paper as a detailed blueprint for building production-grade open assistants.
📺 Source: Two Minute Papers · Published April 07, 2026
🏷️ Format: Deep Dive
