Loopit open-sources Zing-0.5: Building world models with Code + Pixel
新智元 · wechat · 2026-08-26
Loopit (previously praised by Musk) released and open-sourced the real-time interactive world model Zing-0.5, ranking first on the WBench benchmark. It supports "joint control" of spatial exploration and real-time semantic generation, allowing event rewriting within a continuously generated world (e.g., breathing fire while flying, multiple statues opening eyes).
Technical Features:
- Dual Control Paths: Continuous signals are encoded into action residuals for movement, while text conditions rewrite the world via cross-attention; both feed into CausalDiT.
- Anti-Error Training: History perturbations are added during training to force the model to predict stable futures from imperfect contexts, solving consistency issues in long-duration generation.
- Extreme Cost: The 5B model runs >24FPS on RTX 5090, with streaming inference cost approx. $0.01/min.
Architectural View:
Loopit argues that "a world is not a sequence of pixels, but a chain of consequences." Currently, world state, rules, and causality must be maintained by Code, while Pixel handles generative rendering. Zing-1.0 aims to merge AI Coding with video generation.
More from Multimodal
- LichtFeld Studio demonstrates full open-source 3D Gaussian Splatting workflow — janusch_patas · 2026-08-26
- SenseTime open-sources SenseNova U1.5 Lite: 8B unified model with native 4K generation — HeyNayeem · 2026-08-26
- Xiaohongshu FireRedTTS3: Unified Speech Generation and Editing Model — jiqizhixin · 2026-08-26
- Photorealism Has Three Orthogonal Axes and Most Workflows Only Fix Two — 24Latents · 2026-08-26
- Major labels invest $76M in Stability AI — saheedniyi_02 · 2026-08-26
- MiniMax H3-x-Z Text-to-Video Model Trends on HF — joeygambino · 2026-08-26