Skip to content
QbitAI · WeChat

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: Shanghai AI Lab brings agent loops, memory and skill modules into multimodal generation, lifting a 6B model by 14.22 points on average across 5 tasks.

Read the original ↗Export Markdown