🍊 Zing-0.5 — live interactive world model

A 5B causal world model (seedleap/zing-0.5) that hallucinates a world while you walk through it, streaming live as it is generated.

Type a world → press ▶ → steer with W A S D / I J K L (or click the pads).

Your world will appear here.

Describe a world on the right, then press ▶ Enter the world.

READY block 0 0.0 fps 0.0s of world keys —

Move · W A S D

W A S D

Look · I J K L

I J K L

Session

▶  Enter the world ■  End session

Start from an image

🖼️ Drop an initial frame or click to upload

PNG / JPG — becomes the world's first frame. Leave empty for a text-only world.

✕  Clear

World prompt

Rewrite the world mid-session

New prompt

↺  Rewrite now

Takes over from the next block on — the KV cache is kept, so the world morphs instead of cutting.

Settings

Resolution

Autopilot

What the world does until you touch the pad — KEYS FRAMES steps, e.g. W 48, WL 24. - holds still.

Seed

Randomize seed

Sliding-window cache local_attn_size=33, sink_size=5 — the configuration the model card prescribes for real-time rollouts. A session runs for up to 68s of GPU time.

Session recording (24 fps)

Starting worlds

Text worlds — press ▶ once one is loaded

Image-initialized worlds

Prompts and initial frames come from the official zing-world-model examples (Apache-2.0). Live transport: gradio.Server API endpoints (/play, /steer, /stop) plus a raw WebSocket (/ws) streaming binary JPEG blocks.