OpenClaw-style methods reach multimodal generation, with a 6B model beating Nano Banana 2 on some tasks

OpenClaw的风刮到了多模态生成，6B小模型超越Nano Banana 2！

A team led by Shanghai AI Laboratory introduced GEMS, adding Agent Loop, Memory, and Skills to multimodal generation, and reports that 6B Z-Image-Turbo beats Nano Banana 2 on some tasks. The post reports +14.22 average gains on 5 mainstream tasks and +8.92 over the best baseline on 4 downstream tasks; the paper and code are public, but the post does not disclose Nano Banana 2's full setup.

Why it matters: Strong HKR-H/K/R: the hook is a 6B multimodal model beating Nano Banana 2, and the post includes mechanism plus testable deltas (+14.22 / +8.92) with paper and code. It stays below P1 because the article does not disclose the full Nano Banana 2 comparison setup.