Descriptions:
Fahd Mirza tests the InternS2 Preview 397B model from Shanghai AI Laboratory, making the case that Western audiences have largely overlooked one of the most technically ambitious open-source research programs in the world. Shanghai AI Lab has been building the InternLM and InternVL model families since 2021 in collaboration with SensTime and universities including Fudan and CUHK. The new 397B model is explicitly designed for scientific intelligence and long-horizon agentic tasks rather than general leaderboard competition, and is released under Apache 2.0.
Mirza tests the model across three interaction modes — Quick, Task, and Deep Think — starting with a prompt designed to gauge audience-adaptive reasoning: explaining CRISPR mechanics differently for a biology undergrad versus a ten-year-old. He then gives the model a seismogram image and asks it to identify P-wave and S-wave arrival times and generate a self-contained interactive HTML visualization with animated waveform traces and a draggable time cursor.
Benchmark analysis shows InternS2 leads open-source models on MMLU Pro and SimpleQA Verified, and dominates scientific benchmarks including Biology Instructions, MP20 for material structure generation, and Protein Binder 9 for biomolecular interaction design — often by wide margins over both open and closed models. While closed frontier models like GPT-5.5 and Claude Opus 4.8 still edge it out on raw coding and agentic benchmarks like TerminalBench and SWE-Bench Pro, InternS2’s dominance in scientific discovery workflows marks a distinct and underappreciated niche in the open-source model landscape.
📺 Source: Fahd Mirza · Published July 24, 2026
🏷️ Format: Review
