XGEN releases JING, an open-source egocentric interactive world model built on MiniMax-H3
Scobleizer · x · 2026-09-22
XGEN Labs open-sourced XGEN-JING, an egocentric interactive experience model built on MiniMax-H3, available on Hugging Face and GitHub.
Key capabilities:
- Keyboard-controlled navigation through real or imagined worlds via camera controls;
- Interaction and dialogue: steer object manipulation and character conversations with text, generating video and audio jointly;
- Reference conditioning: combine character, object, and scene images to compose experiences and branch actions from the same starting point.
The release includes the four-step bidirectional JING-Flash-v1, inference code, and prompt skills; a causal model and technical report are coming. The demo was validated on six H100 GPUs (one text encoder, one VAE, four DiT with sequence parallelism), with FlashAttention-4 as the default backend.
Related event: XPeng XGEN Open-Sources JING, a First-Person Interactive World Model(3 posts)→
More from Embodied
- This AI-powered laser kills weeds without any chemicals — TinfoilTricorn · 2026-09-22
- Light Origins open-sources Light-O1-Preview, a 6B whole-body robot motion model — CurieuxExplorer · 2026-09-22
- Muse model demoed running on Meta Glasses, tipped as big Connect focus — Scobleizer · 2026-09-22
- VR mocap pipeline captures 1,185 human trajectories to teach Unitree G1 to navigate clutter — Scobleizer · 2026-09-22
- Weekend prototype: colocated spectator view for Pokémon on Snap Spectacles — willeastcott · 2026-09-22
- Auki's HK lab combines IR-SIM and Meta's Habitat for robot navigation in video-reconstructed scenes — broodsugar · 2026-09-22