蚂蚁集团旗下具身AI公司Robbyant开源LingBot-World 2.0,这是一个因果世界模型,支持720p/60fps实时探索。关键突破在于长时间一致性:在20个场景中连续运行60分钟无可见退化。模型采用智能体循环架构,由VLM提议事件、导演智能体持续生成新任务,支持移动、战斗、按需天气及多人共享世界。已发布14B模型权重与代码,论文还描述了可在单消费级GPU部署的1.3B版本。
Open source world models have a fantastic new player!
Interactive world models have had the same flaw for a while now. They look great for a few seconds, then the world quietly falls apart. Textures smear, geometry warps, drift wins.
Robbyant's new LingBot-World 2.0 is the first one I've seen really push past that. It's a causal world model you explore in real time at 720p and 60fps, stress-tested with a single unbroken 60-minute session across 20 scenes with no visible decay. Staying coherent for that long is the actual hard problem, and that's the result here.
The part that got me is the agent loop on top. A VLM proposes events while you play and a director agent keeps seeding new things to do, so the world keeps unfolding instead of sitting still. Movement, combat, weather on demand, even multiplayer in one shared world.