🎮 GameScaling — one 200 ms action chunk, visualised
GameScaling
models are Qwen3.5 VLMs continued-pretrained on 1,600 h of keyboard-and-mouse gameplay.
Given a game screen they answer with exactly one chunk:
<|action_start|>X Y Z ; g1 ; g2 ; g3 ; g4 ; g5 ; g6<|action_end|> —
a mouse displacement in per-mille of the screen, a scroll amount, and the keys held in each of six
33 ms steps. This page parses a reply and draws it on the frame.
What this page is. A static page: there is no model running here.
The Doom reply below is a recorded output shipped by the authors
(examples/doom_battle1_action.txt); you can also paste any reply of your own and watch it be parsed.
Live inference lives in the authors' ZeroGPU Space.