Intern-S2-Preview-397B: scientific AI tests and limits

This review tests Intern-S2-Preview-397B from the InternLM family in its hosted console with scientific explanation and image-to-HTML tasks. The presenter describes it as a multimodal model for scientific intelligence and long-horizon agents, with training across more than 20 scientific domains. He says its weights are available under Apache 2, but cannot install the large model locally.

The quick-mode test asks for gene-editing explanations aimed at a biology undergraduate and a 10-year-old. The presenter praises the technical detail and the simpler analogy, though he acknowledges limited knowledge of the subject and finds the undergraduate response heavy on bullet points. The explanation's scientific accuracy is not independently checked.

In task mode, he uploads a seismogram and requests an interactive HTML visualization with arrival markers and animated traces. Delivery causes trouble: the model produces an unsolicited PDF and refers to local files he cannot access. After follow-up prompts, it supplies HTML as text in four parts, which he combines and opens in a browser. Animation, reset and flag toggles work in his test; the grid toggle fails, and some elements do not move. The seismogram interpretation is not independently validated.

The presenter discusses benchmark charts reporting strong scientific results and mixed coding-agent performance, without reproducing the benchmarks.