NVIDIA's New AI Just Changed Everything

TL;DR

NVIDIA's Nemotron 3 Super is a free, fully open 120B-parameter model that runs up to 7x faster than comparable open models.

Key Points

• 1.Nemotron 3 Super is fully open and free, unlike proprietary rivals. NVIDIA released the model, its 51-page training paper, and the dataset — trained on 25 trillion tokens to produce a 120B parameter assistant matching closed frontier models from ~18 months ago.
• 2.The NVFP4 format makes the model up to 7x faster than similarly capable open models. It compresses math by rounding lower-sensitivity calculations while leaving critical ones intact, achieving massive speed gains with no meaningful accuracy loss.
• 3.Multi-token prediction generates 7 tokens simultaneously instead of one at a time. The system drafts an entire 7-token chunk and verifies it in one pass, delivering another significant speed-up on top of NVFP4.
• 4.Mamba layers solve the memory inefficiency of traditional AI systems. Rather than re-reading all prior context repeatedly, they compress conversation history into selective notes, discarding filler — enabling efficient processing of large amounts of data.
• 5.Stochastic rounding fixes error accumulation caused by compressed arithmetic. Carefully crafted random noise averaging to zero corrects the drift introduced by rounding across hundreds of steps, making the full pipeline accurate and functional.