EVENT DOSSIER
NVIDIA has released Nemotron 3.5 Lightning NVFP4 and Nemotron 3 Ultra…
On August 17, 2026, NVIDIA announced on its developer blog that it had developed the Nemotron 3.5 Lightning NVFP4 model using NVIDIA Model Optimizer, based on QAD (Quantitative Perceptual Training). Previously, on June 26, 2026, NVIDIA had released the Nemotron 3 Ultra NVFP4 checkpoint created using the same tool. Both advancements aim to improve the quantization accuracy and inference efficiency of models through NVIDIA Model Optimizer.
企业 · 官方发布across 2 days
NVIDIA has released Nemotron 3.5 Lightning NVFP4 and Nemotron 3 Ultra…
NVIDIA has released Nemotron 3 Ultra model checkpoints, using the innovative 4-bit floating-point (NVFP4) quantization format to optimize weight transmission efficiency in long context windows.
NVIDIA has released Nemotron 3.5 Lightning NVFP4 checkpoints, developed based on NVIDIA Model Optimizer. While maintaining precision, it increases throughput up to 4 times and compresses the model to 22GB (the original full-precision size is 66GB).
NVIDIA 发布 Nemotron 3 Ultra 模型检查点，采用其创新的 4 位浮点数（NVFP4）量化格式以优化长上下文窗口下的权重传输效率。该 NVFP4 格式是随 NVIDIA Blackwell 架构引入的新技术，旨在将模型权重压缩至更小的数据格式以提升性能。此次 Nemotron 3 Ultra 检查点由 NVIDIA Model Optimizer 工具生成，利用上述量化方案解决大模型在长上下文场景下的高效移动问题。
NVIDIA 发布 Nemotron 3.5 Lightning NVFP4 检查点，在保留精度的同时将吞吐量提升最高 4 倍，并将模型压缩至 22GB（原全精度为 66GB）。该检查点基于 NVIDIA Model Optimizer 开发，旨在帮助开发者针对延迟、速度、内存和计算目标定制 Nemotron 系列模型。