Zing-0.5 renders an interactive video world at 24 frames a second

single source· 1 articles · confidence: medium · first seen 2026-09-14 20:00 UTC

What this means for you

If you build interactive video, the weights and serving stack are released, and a stream-minute is estimated at under a cent of rented compute — cheap enough to prototype against. There is no hosted API and no evaluation date on the scores, so nothing to migrate today.

Zing-0.5 is a 5B-parameter autoregressive world model — a model that generates video of an environment as input arrives, rather than describing one. Keyboard input and typed text instructions are trained together, so a prompt can change events mid-run without restarting generation. It renders 832x480 at 24 frames a second, and the authors estimate about USD 0.009 of server rental per stream-minute. The paper reports 81.0 overall and 88.5 consistency across 158 WBench Navigation cases, with no evaluation date given. Weights, inference code and a serving implementation are released.

Key facts

  • ·Zing-0.5 is a 5B-parameter autoregressive world model with joint keyboard and text control. source
  • ·It renders 832x480 at 24 FPS using four-step generation. source
  • ·The authors estimate a server rental cost of approximately USD 0.009 per stream-minute. source
  • ·It scores 81.0 overall and 88.5 on consistency across 158 WBench Navigation cases; the paper gives no evaluation date. source
  • ·Model weights, inference code and the Zing-SGLang serving implementation are released. source
  • ·Text instructions are temporally aligned with keyboard input and jointly annotated videos during training. source

What the sources say

Sources

The original reporting. Follow these — they did the work.

← the wire