Some of NVIDIA's models also tend to have interesting architectures. For example, usage of the MAMBA architecture instead of purely transformers: https://developer.nvidia.com/blog/inside-nvidia-nemotron-3-t...