Intern-S2-Preview: Smaller AI for Scientific Discovery

InternLM has released Intern-S2-Preview, a new 35-billion parameter scientific multimodal foundation model. This model is challenging the idea that scientific AI needs trillion-scale models. What's interesting is its performance. The team states Intern-S2-Preview offers comparable results to its larger, trillion-scale Intern-S1-Pro model on several professional scientific tasks. This is significant because scientific AI currently faces high costs. Startups in fields like materials discovery or protein analysis need models that can handle scientific data and tools, but also keep inference costs manageable. Intern-S2-Preview aims to fill this gap. It uses a method called task scaling, focusing on expanding the difficulty and diversity of scientific tasks during training. This approach suggests that more rigorous scientific training can achieve similar results to simply increasing model size. While still a 35-billion parameter model, it offers specialized features like stronger spatial modeling for small-molecule structures and material crystal structure generation. The bottom line is this model could offer a more efficient and cost-effective way for scientific startups to integrate AI into their workflows.