Intern-S2-Preview

Next-generation flagship for Shanghai AI Lab's scientific multimodal foundation-model line. ~35–36B total / ~3B active hybrid-attention MoE with 262K context, continued-pretrained from Qwen3.5 with domain extensions for materials, biology, time-series, and crystal-structure generation.

Reported benchmarks: SWE-bench Resolved 64, MMLU-Pro 88, MathArena HMMT Feb 2026 87.31 — competitive with the much larger Intern-S1-Pro at a fraction of the active-parameter count.