Jev 驱动的实时 Doom 演示
发布线程展示游戏场景中的实时决策效果。Diogo 说明演示约每秒调用 10 次 Jev,并附有 2 分 43 秒视频。

作者训练了一个0.8B的多模态决策模型,只通过屏幕画面即可生成键盘鼠标操作,在MacBook本地以5 FPS运行玩Minecraft。
I trained a jev like multimodal decision model to play Minecraft. It only sees the screen. You enter a prompt, it outputs keyboard and mouse keys. The 0.8B model is running locally on my macbook at 5 FPS. Vision is all you need 👇