Nemotron 3 Super: Open-Source AI Assistant Outruns Closed Models

Nemotron 3 Super arrives as a free, open‑source AI assistant. Its launch is accompanied by a 51‑page research paper that spells out the training process, the dataset of 25 trillion tokens, and a 120‑billion‑parameter architecture. The model’s performance roughly matches closed‑frontier systems that were state‑of‑the‑art about 18 months ago. As the host notes, “This is an AI assistant that is free for all of us forever, but not just the model itself.”

Performance and Efficiency Metrics

Two versions of the model are offered: a standard BF16 format and a highly optimized NVFP4 format. The NVFP4 variant runs about 3.5 times faster than its BF16 counterpart and can be up to 7 times faster than other open‑source models with similar capabilities. Despite the speed boost, accuracy stays comparable across both versions, fulfilling the claim that “the story is not just the similarly smart part, the story is that it is 7 times faster while it is similarly smart.”

Technical Innovations

NVFP4 Compression

NVFP4 works by rounding off digits in mathematical operations that are not critical to the final result, thereby cutting the computational workload. Sensitive calculations remain exact, and stochastic rounding injects carefully crafted zero‑mean noise to stop rounding errors from compounding over long sequences.

Multi‑token Prediction

Instead of generating one token at a time, the system predicts and verifies seven tokens simultaneously. This parallelism reduces the number of inference steps required for a given output, directly contributing to the observed speed gains.

Mamba Layers

Mamba layers adopt a memory‑efficient approach: the model reads the input once, takes highly compressed notes, and discards filler words. As one quote puts it, “Memory is precious. So instead, read the book only once, and take highly compressed notes.”

Stochastic Rounding

Stochastic rounding adds random noise that averages to zero, preventing the gradual drift that can occur when deterministic rounding is applied repeatedly. This technique safeguards the model’s accuracy during the multi‑token prediction process.

Industry Implications

The open release of Nemotron 3 Super signals a potential shift away from the dominance of closed, proprietary AI systems. NVIDIA is reportedly investing tens of billions of dollars into fully open AI initiatives, suggesting a strategic move toward transparency and community‑driven development. The speaker emphasizes, “They spilled all the secrets,” highlighting the unprecedented level of technical disclosure. If Jensen at NVIDIA is indeed committing such resources, the competitive landscape may soon favor open‑source models that combine high performance with rapid inference, challenging the long‑standing advantage of closed‑source offerings.