GLM-5.3 物理模拟测试战胜隐身模型 Space Bunny Alpha
有人拿牛顿摆实测了 GLM-5.3 和 OpenRouter 上的神秘新模型，物理模拟谁更靠谱一看便知，还有难倒两个模型的翻车场景。
@aimlapi 用牛顿摆做了 5 场景物理模拟对比测试，GLM-5.3 击败了当天新上线的隐身模型 Space Bunny Alpha。Space Bunny Alpha 目前在 OpenRouter 和 OpenCode 上线，支持 100 万 token 上下文窗口和原生多模态输入，代码能力较强。在龙卷风和水滴两个更难的场景中，两个模型表现都不理想。测试结论指出，碰撞时机和动量传递一旦出错，靠增加环境细节无法弥补。
GLM-5.3 beat Space Bunny Alpha (the new stealth model launched today) on Newton’s cradle in a 5-scene physics test done by @aimlapi .
Space Bunny Alpha is currently stealth, launched on OpenRouter/OpenCode, with a 1-mn token context window, native multimodal input, strong coding performance.
The tornado and water drop were even harder. Neither model handled those well.
visual result is constrained by the dynamics. you can't compensate for incorrect collision timing or momentum transfer by adding more detail to the environment.
