NCP-ArchPreview Technical Report: Moving towards Latent Space Language Models through Next Concept Prediction
An 8.9B-parameter latent-space language model using next-concept prediction matches OLMo-3-7B pretraining loss with only 51.3% of the training tokens.