Lingbot World 2 — realtime interactive world model
A navigable world generated chunk-by-chunk at 720p/60fps for over an hour, with NPC agents and prompt-driven events — an open world model for embodied AI.
🌍 World Models
Tracking generative "world models" — models that stream navigable, interactive 3D worlds in real time. Fast-emerging and relevant to embodied AI, simulation, and spatial capture. Reference only; verify specifics before acting.
Landscape
Genie (Google DeepMind)
Interactive world model that generates playable environments from an image.
Genie 3 announcement (DeepMind) · checked 2026-09-29
Oasis (Decart)
Realtime interactive game world streamed token-by-token at playable frame rates.
MIT Technology Review on Oasis (2024-10-31) · checked 2026-09-29
GameNGen (Google)
A neural game engine that simulates a playable world end-to-end.
Diffusion Models Are Real-Time Game Engines (ICLR 2025) · checked 2026-09-29
NVIDIA Cosmos
World-foundation models for physical AI, used alongside robotics and simulation stacks.
NVIDIA Cosmos (product page) · checked 2026-09-29
World Labs (Fei-Fei Li)
"Spatial intelligence" — generating 3D worlds from single images.
Atlas: A World Model for Spatial Intelligence (World Labs) · checked 2026-09-29
Odyssey
Cinematic world models for film and video production.
Introducing Odyssey-2 (Odyssey) · checked 2026-09-29
Lingbot World 2
Open-source realtime world model (720p/60fps, NPCs) — from the Modern Research feed below.
LingBot-World 2.0 (Robbyant) · checked 2026-09-29
Landscape v1 · published 2026-09-29 · curated 2026-09 · 7/7 entries link the source they were checked against. Hand-curated, not a live feed: check a primary source before relying on a capability or a date.
Modern Research
Realtime interactive world generators are arriving
A new class of 'world models' streams navigable, interactive 3D worlds in real time (720p/60fps) from prompts or key presses — alongside Genie, Oasis, and Cosmos. The emerging space to watch for embodied AI and simulation.
DSec — the training-environment layer, not the GPU, is the bottleneck
DeepSeek open-sourced DSec, the sandbox infrastructure behind its agentic RL: ~3M sandboxes/day, ~380K concurrent, four isolation backends from FnCall to full VM, and on-demand image loading that starts an 8,192-container burst in ~35 min vs >60 min the Docker way. The load-bearing claim for world models: because every new agent generation needs new tasks to practise on, the constraint on self-improvement is the supply of environments rather than compute. Section 6.1 answers it — agents build environments on the training infrastructure itself, and pack_diff turns any interactive session into a reusable training ground.
SpatialWorld 1.5 — an image plus a camera path becomes a walkthrough
Give it one image and a camera trajectory and it renders video walking through that scene. Multiple images, a panorama, or a source video also work, and a video input carries its own motion along the path; frames can be frozen mid-path for bullet-time. Open code, built on Wan 2.1 1.3B, under 6 GB — consumer-GPU sized. Relevant here: the cheapest route from the property stills to a walkthrough clip.