Shanghai AI Lab and SJTU unveil NCP-ArchPreview, the first 8.9B base model using Next Concept Prediction. Pre-trained on 5.73T tokens, it uses only 51.3% of the token budget to match OLMo-3-7B's loss, boosting convergence by 1.95 times and computational efficiency by 1.74 times. This latent space approach improves downstream performance, including a 6-point gain in GSM8K. Fine-tuning requires just 17M parameters, outperforming LoRA without forgetting, while speculative decoding sees acceptance lengths increase by up to 7.59% in code tasks, offering a new path for inference acceleration.