Inkling is a Mixture-of-Experts transformer with 975B total parameters and 41B active. It supports a context window of up to 1M tokens. Pretraining covered 45 trillion tokens of text, images, audio, ...