NVIDIA AI has expanded its Megatron Core framework to include support for advanced optimizers like Muon, alongside research optimizers such as MOP and REKLS. This development aims to improve the efficiency of training large-scale models, including those at the scale of Kimi K2 and Qwen3 30B. The announcement highlights the need for techniques beyond standard data parallel methods to achieve efficient model training. However, specific performance metrics or implementation details have not been provided.