HKU and Tencent open-source Multics, a multi-player world model that trains on 4 players but scales to 10
Scobleizer · x · 2026-10-14
Researchers from the University of Hong Kong and Tencent's ARC Lab (IEG) open-sourced Multics, a generative multi-player world model that generates videos of a shared environment where each player follows their own controls while overlapping views stay consistent.
Key points:
- Existing approaches trade off: fully-connected communication costs quadratic attention as players grow; centralized communication mixes all features, causing interference and capacity competition.
- Multics uses hierarchical communication (player → group → global) with shared group-token parameters and dynamic grouping based on camera trajectories.
- This enables training on 4 players and inference on 6, 8, or 10 players without retraining, at lower self-attention cost, with robustness to interference between unrelated events.
Inputs per player include initial view, text prompt, actions, and camera trajectory. The project page includes demos and BibTeX.
Related event: HKU and Tencent open-source Multics, a multi-player world model(2 posts)→
More from Multimodal
- Oscilloscope Diffusion restyles audio-reactive visuals with diffusion models, open-sourcing soon — uisato · 2026-10-14
- HeyGen Voice tops Artificial Analysis TTS blind-test leaderboard, API 50% off — HeyGen · 2026-10-14
- Scenario open-sources a full 3D pipeline of Blender agents, driven by Claude Code — DavidmComfort · 2026-10-14
- Local models fall short of nano banana 2 for multi-reference face scene generation — classified18 · 2026-10-14
- RVM: Likelihood-Free Reward-Based Velocity Matching for Scaling Diffusion RL — gowthami_s · 2026-10-14
- Everygen Wizard Launches: Full Video Workflow in One Conversation — Scobleizer · 2026-10-14