Many users are excited about LingBot-World 2.0 because its 1.3B open weights enable coherent 720p 60 FPS world simulation on consumer GPUs like the 3060.
Based on 4 visible X reactions from 7 accounts; directional sample.
Ask a question below.
Published answers will appear here.
@omarsar0 1.3B means this runs on my 3060. Coherent for a full hour at 720p 60fps is genuinely impressive for an open weight world model.
@omarsar0 the 1.3B running on consumer GPU is what makes this actually exciting real world models arent just server flexes anymore
@omarsar0 that is actually insane for a consumer gpu
🧵 5. The agentic harness keeps the generated world moving without a person scripting every moment. Two agents underneath: a pilot agent plans and executes character behavior; a director agent seeds fresh content and events so the world doesn't run dry. Plus two interaction modes — Mode A (direct semantic, no masks) and Mode B (SAM-tracked object interaction) — and world intervention like day↔night, weather, or spawning entities.
@robbyant_brain's newly released LingBot-World 2.0 💻 Github: https://github.com/Robbyant/lingbot-world-v2 🎮 Try it online: https://reactor.inc/lingbot-world-v2 📄 Paper: https://arxiv.org/pdf/2607.07534 🌐 Watch the videos in action: https://technology.robbyant.com/lingbot-world-v2 🤗Weights in Huggingface: https://huggingface.co/collections/robbyant/lingbot-world-v2
@robbyant_brain's newly released LingBot-World 2.0 💻 Github: https://github.com/Robbyant/lingbot-world-v2 🎮 Try it online: https://reactor.inc/lingbot-world-v2 📄 Paper: https://arxiv.org/pdf/2607.07534 🌐 Watch the videos in action: https://technology.robbyant.com/lingbot-world-v2 🤗Weights in Huggingface: https://huggingface.co/collections/robbyant/lingbot-world-v2
🧵 6. Interaction goes far beyond camera movement: WASD and IJKL handle navigation, while hotkeys trigger combat, archery, spell-casting, shooting, weather changes, and user-registered events. The interface also supports multiple players inside the same generated world. https://x.com/rohanpaul_ai/status/2075768187245990107/photo/1
🧵 3. Consistency distillation makes the model fast, but speed means little if its own mistakes keep piling up. So LingBot-World 2.0 trains on long sequences generated by the model itself, not only clean examples. It gets practice handling the imperfect history it will actually face while someone is playing.
🧵 7. The honest limitation is memory: visual stability can last a long time, but a region that leaves the context window may be generated again rather than remembered exactly. The paper describes it as persistent in appearance, not in identity. https://x.com/rohanpaul_ai/status/2075768191566147807/photo/1
🧵 8. Watch LingBot-World 2.0 in action. @robbyant_brain's newly released LingBot-World 2.0 🤗 Weights: https://huggingface.co/collections/robbyant/lingbot-world-v2 https://x.com/rohanpaul_ai/status/2075768196511252609/video/1
🧵 9. Some more results of LingBot-World 2.0 world model. https://x.com/rohanpaul_ai/status/2075768200705572972/video/1
Many users are excited about LingBot-World 2.0 because its 1.3B open weights enable coherent 720p 60 FPS world simulation on consumer GPUs like the 3060.
Based on 4 visible X reactions from 7 accounts; directional sample.
Ask a question below.
Published answers will appear here.
🧵 6. Interaction goes far beyond camera movement: WASD and IJKL handle navigation, while hotkeys trigger combat, archery, spell-casting, shooting, weather changes, and user-registered events. The interface also supports multiple players inside the same generated world. https://x.com/rohanpaul_ai/status/2075768187245990107/photo/1