Multics Scales Multiplayer World Models

Multics tackles a real bottleneck in interactive video generation: unrelated players should not compete for the same global representation. The architecture is promising, but its enormous GPU requirements make it a research prototype rather than a practical multiplayer engine today.

• –Dynamic grouping based on camera trajectories limits unnecessary cross-player attention.
• –Shared group-token parameters enable scaling beyond the number of players seen during training.
• –The system reports stronger cross-player consistency, controllability, and interference robustness than centralized or fully connected baselines.
• –Open-sourced Apache-2.0 code makes the work unusually reproducible, though model weights and example content retain upstream terms.
• –Inference reportedly requires roughly 60 GiB of GPU memory for four players and more than 106 GiB for ten, sharply limiting experimentation.