Shanghai AI Lab and SJTU LUMIA Lab released the NCP-ArchPreview Tech Report, proposing the Next Concept Prediction pre-training task. The first 8.9B-parameter discrete latent-space foundation model matched OLMo-3-7B's final pre-training loss using only 51.3% of the token budget, achieved a 1.95x equivalent convergence speedup and a nearly 6-point GSM8K improvement, with all assets open-sourced.
Shanghai Artificial Intelligence Laboratory, together with Fudan University, the Institute of Software of the Chinese Academy of Sciences, East China Normal University and others, has developed the Shusheng Xinghe agent model. Based on the AI4AI paradigm with over 400 million parameters, the model can predict global weather for the coming week within one minute. The team conducted an empirical analysis of 769 task records from 56 researchers, validating a human-in-the-loop collaboration model where AI executes and humans decide, and explored a controllable recursive self-improvement (RSI) paradigm.
On September 14, Shanghai Artificial Intelligence Laboratory open-sourced the ATRIA Dawn Preview model. Trained on over 100GB of global weather data with over 0.4B parameters, the model can predict global weather one week ahead within one minute, and outperforms FourCastNet on some metrics. Muxi completed Day0 adaptation for the model on the same day.