According to CoinMeta and OneMillion.AI, Shanghai AI Lab and Shanghai Jiao Tong University have released a model with 8.9 billion parameters, ncp-archpreview. This model is not only trained to "guess the next token" but also predicts the corresponding internal conceptual signals for a short segment following it, to assist in the generation by token. In this way, the model can anticipate the subsequent content in advance, thereby improving generation efficiency. In four tests, the average number of correct predictions per validation by the large model increased from 5.933 to 6.180, representing a 4.17% improvement. In the humaneval test, the improvement was 7.59%. After incorporating these signals, the draft model only required an additional 40,000 parameters, and using about 85% of the training computational resources of a regular transformer, it was able to reduce training errors to a level that is nearly comparable.