Kyutai Labs开源了MIRA，一个在神经网络里直接跑《火箭联盟》的模型，5B参数、20fps，不用游戏引擎纯靠学习生成，比传统游戏模拟器有意思多了。
Kyutai Labs与Gen Intuition联合Epic Games发布MIRA模型，参数规模5B，在单张Nvidia B200上以20帧/秒实时生成完整2v2火箭联盟比赛画面，无物理引擎或渲染引擎。模型仅从视频和控制器输入自监督学习，已开源代码、技术报告及1000小时数据集。局限在于模型记忆仅约4秒，回放时会遗忘比赛目标并幻觉生成替代画面。
Four people just played a full 2v2 Rocket League match inside a neural network. Every frame they sa...
Four people just played a full 2v2 Rocket League match inside a neural network. Every frame they saw came out of a 5B-parameter model dreaming the whole game, 20 frames a second, on a single Nvidia B200, with "no physics engine, no rendering engine, and no explicit 3D representation at all," in the team's words. It learned Rocket League from video and controller inputs alone. @kyutai_labs and @gen_intuition built MIRA with Epic Games, and open-sourced the code, the technical report, and a 1,000-match-hour dataset today. A funny note from the limitations section: the model remembers only around four seconds... So when a replay rolls, it has already forgotten the goal and just hallucinates one instead. Your browser does not support the video tag. 🔗 View on Twitter 💬 3 🔄 0 ❤️ 12 👀 2698 📊 4 ⚡