Game World Field Guide

A foundational world model developed by Google DeepMind, first opened to the public on January 29, 2026 as the 'Project Genie' experimental prototype (for Google AI Ultra subscribers). Genie 3 generates interactive virtual environments in real time at 720p resolution and 24FPS; after the March 2026 update, stable duration improved from a few seconds to about one minute. Supports text/image input to generate worlds, WorldRemixing for secondary creation, and prompt-triggered world events. At Google I/O in May 2026 it showcased deep Gemini integration and was defined as 'one of the most advanced world models'

A spatial intelligence company founded in April 2024 by 'AI godmother' Fei-Fei Li. In February 2026 it closed $1 billion in funding (Autodesk led with $200M, NVIDIA/AMD/a16z participating), at a $5 billion valuation — a 5× jump in under two years. Core product Marble: generate persistent, editable 3D worlds from a single image or text, emphasizing spatial structure and geometric consistency. On April 15, 2026 it open-sourced the Spark 2.0 rendering engine — any phone/browser can smoothly render 3D scenes with over 100 million Gaussian points. Partnering with Autodesk to bring world models into 3D creation workflows (Unreal/Unity/Blender)

An advanced machine intelligence lab founded in Paris in January 2026 by Turing Award winner Yann LeCun. In March it closed a $1.03 billion seed round (NVIDIA/Samsung/Bezos Expeditions/Temasek/Cathay Innovation participating) at a $3.5 billion pre-money valuation, setting the record for the largest seed round in European AI startup history. CEO is former Meta FAIR engineering lead Alexandre LeBrun; chief scientist is Xie Saining (who twice turned down OpenAI). Core route JEPA (Joint Embedding Predictive Architecture): rather than generating pixels, it predicts world states in an abstract representation space — LeCun calls it 'the only way to AGI'. Pursuing an open-source route while maintaining collaboration with Meta

A real-time AI world generation model from Israeli AI startup Decart (founded by IDF 8200 Unit veterans). In November 2024 Oasis 1.0 launched — the world's first interactive game (Minecraft-like) generated frame-by-frame in real time by AI with no game engine. In May 2026 it closed $300M (cumulative over $450M) at a valuation near $4 billion (NVIDIA/Adobe/Toyota/eBay/Karpathy participating). On June 10, 2026 it released Oasis 3 — real-time generation of realistic driving environments, priced at $0.02/second, with an open API. The CEO's goal: to build 'the first world model developers can program on top of'

A multimodal world model released and open-sourced by Tencent on April 16, 2026. It can automatically generate, reconstruct and simulate 3D worlds from text/image/video input, directly exporting 3D asset files editable for secondary creation (Mesh/3DGS/point cloud). The technical architecture is 3D-centric, building a complete closed loop from sparse input to an interactive 3D world via modules like HY-Pano 2.0/WorldNav/HY-WorldStereo/HY-WorldMirror 2.0. It imports seamlessly into Unreal/Unity game engines. On the same day Alibaba released Happy Oyster in a 'head-to-head'. Tencent Games has already achieved AI NPC closed loops in products such as Game for Peace

A world model open-sourced on January 29, 2026 by Ant Group's LingBot Technology. The core LingBot-World-Base relies on a scalable data engine, learning physics and causal relationships from large-scale game environments to achieve high-fidelity, interactive simulated world generation. It leads in video quality, dynamic performance, long-temporal consistency and interactivity. Through multi-stage training and parallel inference optimization, it provides a 'digital playground' for embodied intelligence, autonomous driving and game development

An iterated world model released on April 10, 2026 by AGIBOT (智元机器人). Core breakthroughs: minute-level long-temporal reasoning + multi-view 3D consistency + a built-in General Reward Model, for the first time embedding a reinforcement-learning closed loop inside the model world, achieving 'RL in WM'. It builds an interactive, trainable, evaluable 'physical evolution engine', supporting robots in completing large-scale closed-loop training within simulated environments

NVIDIA's world foundation model platform for physical AI, highlighted at GDC 2026. Cosmos provides pre-trained World Foundation Models on which developers can fine-tune domain-specific simulation environments (autonomous driving/robotics/industrial). Gen-3C belongs to the 'spatial-native / 3D-guided' school, alongside World Labs Marble as a representative of that route. NVIDIA also provides digital twin and robot simulation infrastructure through Omniverse/Isaac Sim; companies like Decart and AGIBOT are ecosystem partners

An AI game character engine company founded by former Google employees, with cumulative funding over $120M (Microsoft/Disney/Meta/Intel/Samsung/Lightspeed participating) at a $500M valuation. Core product: rapidly generate 'souled' NPCs with memory/background/emotion/EQ/body language via natural-language description. In late 2023 it signed a partnership with Microsoft Xbox (character engine + Copilot assistant), and collaborates with giants like NetEase/Disney. By 2026 AI NPCs have become a standard direction in the game industry, with domestic solutions such as Tencent Hunyuan's role-playing model rising in parallel

A one-stop AI game creation platform developed by Hangzhou Jiyi Artificial Intelligence Technology, built on a self-developed game-industry-specialized large model, enabling natural-language-driven full-pipeline development (planning → art → code → audio → deployment). Core capabilities: upload a video to generate a playable game, one-click generation of game characters with full skeletal animation, and a built-in professional game numerical engine SOON-FX. The second beta launched on March 31, 2026, winning the '2026 AI Business Case Excellence Award' and 'China AI New Force TOP30'. Positioned for the 'everyone-creates era', letting individual developers make commercializable, high-quality games

A browser-based zero-code/low-code AI game creation platform that rapidly generates 2D/3D game prototypes from natural-language descriptions for instant online play. Supports visual asset generation, game logic code generation and real-time iteration. Ideal for game jams, education and rapid prototyping. Visuals lean minimalist; it does not support project file export, and complex AI / networking / multi-platform sync capabilities are limited. In 2026 it was rated a representative of the second tier of AI game creation tools

The most incredible funding story in the 2026 AI gaming track: 31-year-old Dutch founder Pim de Witte raised $320M at a $2.3 billion valuation (Khosla Ventures/Jeff Bezos/Eric Schmidt/DeepMind researchers participating). Core method: train AI on game recordings from titles like Fortnite to learn physical reasoning, spatial intuition and causal understanding — not training 'an AI that plays games', but using game data to train 'an AI that understands the world'. Goal: transfer the physical intuition learned in games to robotics, autonomous driving and industrial control

Alibaba's open interactive world model released June 17, 2026; registration is open at happyoyster.cn. Unlike text-to-video's one-way text→video mapping, it learns state-transition rules (current state/action → next state/action), reasoning through action-to-feedback causal chains while keeping long-range consistency. Two core modes: Adventure (generate a deeply explorable digital world from one sentence or one image, with forward/dash/crouch/jump/attack controls, vehicle driving and weapon combat, and up to 1 minute of continuous real-time movement) and Directing (a god view with pause, rewind and branching narrative, rendering 3+ minutes of real-time 480p/720p footage). Joint work with Nanjing University on a world-model benchmark; API to open soon