Zing-0.5 ships open weights for a real-time playable world model
A 5B model, 832×480 at 24fps for ~$0.009 a stream-minute, serving stack included — as open playable worlds race ahead and Git becomes shared agent memory.
Copy markdown
Weights, inference code, and a serving stack — all out
Zing-0.5 is a 5B autoregressive world model released with weights on Hugging Face, inference code, and a Zing-SGLang serving stack. It runs 832×480 video at 24fps for roughly $0.009 per stream-minute — a live, controllable world you can host without a research budget.
You steer it live, with a keyboard and typed prompts
Unlike a video generator, Zing lets you navigate a generated world and change what happens mid-stream through joint keyboard and online text control. It scores 81.0 overall and 88.5 on consistency across 158 WBench navigation cases — enough for a playable prototype or an agent training environment this weekend.
The open playable-worlds race is suddenly crowded
Zing lands beside Skywork's fully open Matrix-Game 3.0 (720p at 40fps, 5B, minute-long memory, code + weights + report) and Shanda's Apache-2.0 AlayaWorld (15B, 720p, 60s+, weights still 'coming'). Together they're closing the gap to closed systems like DeepMind's limited-access Genie 3.
Elsewhere: Git as shared memory for agent swarms
Agora stores collective AI research as an append-only DAG in Git, so every claim is a commit anyone can check out and rerun — no duplicated work across sessions. In a 12-day run, 13 agents posted 1,703 contributions and closed 62% of the gap to a trained GPT-2 124M on a weight-transfer task.