Zing-0.5: 5B Real-Time World Model With Joint Keyboard-Text Control, $0.009 per Stream-Minute
Mingyang Chen · hf · 2026-09-17
Zing-0.5 is a 5B autoregressive world model built for playability: users explore generated worlds and steer events via joint keyboard and online text control.
Three technical contributions
- Unified action and text conditioning: magnitude-aware keyboard inputs combined with temporally aligned text instructions and jointly annotated videos, learning navigation and event control in the same sequence
- Event-scale supervision: a segment-level teacher trained on connected multi-prompt videos supervises a block-level causal student via distribution-matching distillation
- Low-cost real-time interaction: four-step generation with context-preserving streaming supports 832x480 inference at 24 FPS, at an estimated $0.009 per stream-minute of server rental
It scores 81.0 overall and 88.5 consistency across 158 WBench Navigation cases, and demos a text-directed event change mid-navigation without restarting generation. Model weights, inference code, and a Zing-SGLang serving implementation are open-sourced.
More from Multimodal
- ComfyUI Trellis2 and Pixal3D meshes keep coming out with random holes — nirurin · 2026-09-17
- Epic's Tim Sweeney: world models can make movies locally playable — smtabatabaie · 2026-09-17
- YuE2: the 3B open-source music model punching above its weight — renderartist · 2026-09-17
- Runway-made looping video reimagines Ganesh Chaturthi's made-new-not-made-new theme — CurieuxExplorer · 2026-09-17
- Use Timer Nodes When Benchmarking Generation Speed, Console Times Lie — CurrentMine1423 · 2026-09-17
- Creative short film made with Seedance 2.5 shows cinematic crowd-freeze scene — SimplyAnnisa · 2026-09-17