Shanghai AI Laboratory has open-sourced the foundational large model Shusen-S2. Its general capabilities rank among the top open-source models, and in areas such as biology, materials science, chemistry, and various scientific long-term tasks, it can already compete with leading closed-source models.

The architecture is the most notable aspect this time. Shusen-S2 adopts a Memory Decoder architecture, introducing a plug-and-play specialized memory module — effectively giving the model an "external brain" that can be swapped at any time: when deep expertise in a specific field is needed, the corresponding memory can be attached without retraining the entire model, while still maintaining its original general capabilities. This design makes domain expansion and generality no longer conflicting for the first time.

In addition to the memory mechanism, Shusen-S2 also enhances long-range reasoning and agent execution capabilities. For research scenarios, these two aspects are the most challenging — a scientific problem often involves many steps and repeated use of tools, and the model must think deeply and act correctly. Shusen-S2 strengthens both aspects, clearly indicating its goal: to make open-source models truly capable of handling serious scientific research tasks, not just scoring high on general benchmarks.