World models guide: Genie 3, GWM and open weights, updated weekly

Updated 9 Nov 2026 · next update 16 Nov 2026

Key facts as of 26 Sep 2026

• What one is: a world model is an AI system that simulates aspects of the world so an agent can predict both how an environment will develop and how its actions change it, which is how Google DeepMind describes the category. (source)
• Genie 3: DeepMind's interactive world model generates a navigable environment at 720p and 24 frames per second from a text prompt and holds consistency for a few minutes, with its own limits listed as a small action space, weak multi-agent interaction and short sessions. (source)
• Runway's GWM Worlds 2: shown on 3 September 2026 as a research preview, it generates continuous 720p at 24 fps with 48,000 Hz audio, is steered by text actions and camera motion, and has no preset session length. (source)
• Runway Characters: built on Runway's GWM-1 general world model, it animates a single reference image at 24 fps with an effective 37 ms of model time per frame and 1.75 seconds of server-side turnaround from the end of speech to the first response frame. (source)
• PixVerse R2: announced on 22 September 2026, it makes input persist so an action shapes what happens later in the same session, and PixVerse says R1, launched in January 2026, was the first real-time world model. (source)
• World Labs: a spatial intelligence company whose first product, Marble, generates persistent 3D worlds from text, images, video or 360 panoramas, with an API listed alongside it. (source)
• Odyssey: raised a $310 million Series B, describes its models as foundation world models and names Agora-2 among them, and its team comes from DeepMind, Waymo and Tesla; its API documentation offers Odyssey-2 Pro through real-time interactive streams. (source, API docs)
• Open weights: Astronex-World 1.0, submitted to arXiv on 17 September 2026, is a 5B model built on the Wan2.2-TI2V-5B prior that streams 832x480 video at 24 fps on one Nvidia L20 48GB GPU and scores 73.5 on WBench Navi and 70.0 on WBench Full. (source)
• Hardware: AWS's Neuron Science team and Reactor generated real-time video above 16 frames per second on a single Trainium2 chip, with one kernel dropping from five seconds to 1.8 milliseconds. (source)

The frontier of world models is now split between the closed research previews and the open research: Google's Genie 4 runs real-time audio with prompt-steered simulation behind a waitlist, while KAIST's L3G-World and Tencent's Hy-Nomas publish their methods and weights with their authors' own numbers and no independent scores. (source)

What a world model is

A world model is an AI system that predicts how an environment behaves. Google DeepMind's Genie 3 announcement puts it plainly: world models use an understanding of the world to simulate aspects of it, so agents can predict both how an environment will evolve and how their actions will affect it. That makes them both a product category and a piece of research infrastructure, because a world model is a place to train agents without a physical world.

The distinction that matters commercially is between generating a clip and running a world. A video model returns a fixed result: change something and you prompt again. A world model keeps generating while you are inside it, takes new input mid-session and holds its state. PixVerse's R2 announcement describes the difference from the inside: in earlier real-time generation an input affected the current moment and then passed, and in R2 an input updates the running state of the world and shapes what follows.

Three things have to be true at once for this to work, and they explain why the field is hard. The world has to stay consistent over time, the model has to run fast enough that a late frame does not break the experience, and the session has to persist. Runway's engineering post on Characters puts the frame budget at 42 milliseconds at 24 fps, before audio processing and network transit.

What world models cost

Costs split by whether you are running one, calling one, or building on one, and most of the field is not yet purchasable.

• Research previews: Genie 3 and Runway's GWM Worlds 2 are shown rather than sold. Genie 3's announcement describes it as a limited research preview with early access for a small cohort of academics and creators, GWM Worlds 2 carries a research-preview label on Runway's own page, and neither publishes a price. (Genie 3, GWM Worlds 2)
• APIs: Runway Characters is available through the Runway API and the Runway web and mobile apps, and Odyssey's API documentation offers Odyssey-2 Pro, a general-purpose world model, through interactive streams for real-time generation as well as viewable streams and batch simulations, with a developer account and API key. Both can be called today. (Runway, Odyssey)
• Open weights: Astronex-World 1.0 is free to download under Apache 2.0 from the authors' release, and the model streams on one Nvidia L20 48GB GPU, so the cost is hardware and time rather than a rate card. (arXiv)
• 3D output: World Labs' Marble generates and exports 3D worlds rather than a video stream, and the company lists an API beside the product without publishing a price. (World Labs)
• Infrastructure: serving one at scale is a separate problem. Reactor's chief technology officer, Bryce Schmidtchen, told Amazon Science: "It's about real time, it's about low latency, and it's about doing that as efficiently as you can at scale." The AWS and Reactor work on Trainium is published research, not a priced product. (Amazon Science)

Previews with no published price are the normal state here, so any comparison of what a world model costs is really a comparison of access: an API you can call today, a preview with limited access, or weights you can run.

What world models do

They simulate rather than render. Genie 3 builds a navigable world frame by frame from a description and the user's actions, and DeepMind notes the consistency is emergent rather than the result of an explicit 3D representation like a NeRF or Gaussian splat. Actions include movement plus text-based events, which DeepMind calls promptable world events.

They keep state. PixVerse's R2 carries input forward, so a choice made early changes what happens later, and its example is a story encounter that branches live depending on what a player offers. Runway's GWM Worlds 2 turns the same idea into a directorial mode: a text action can be addressed to a subject or to the scene, so a campfire flares and a sunset fades to night in the same running world.

They generate sound as well as picture. GWM Worlds 2 generates audio at 48,000 Hz alongside the video, and a text action can be a spoken reply as well as a movement. (source)

They drive characters. Runway's Characters animates a single reference image at 24 fps with lip sync and head motion, with no fine-tuning, and 1.75 seconds of server-side turnaround, and Runway says Characters can join Zoom, Google Meet or Teams calls and respond in real time. (source)

They need the compute fixed. AWS and Reactor's work is a window into the real constraint: the teams used Rolling Forcing as a stand-in for autoregressive diffusion video, and the kernel-level rewrites took one operation from five seconds to 1.8 milliseconds and cache copies from 23 to 1.9 milliseconds per layer, to sustain above 16 frames per second on a single Trainium2 chip.

They are being made in 3D too. World Labs' Marble generates spatially consistent 3D worlds from text, images, video or panoramas and exports them in 2D and 3D formats, which is a different output shape from a streamed video world and a different set of customers.

Who world models are for

Today, four buyers. Researchers are the first audience, and they are the one group the previews actually serve: Genie 3's announcement frames the work around training agents, and the model was tested with DeepMind's SIMA agent navigating generated worlds. Robotics and autonomous driving are the second: Odyssey names both among the fields it wants world models to change, and Runway lists robotics and embodied-agent simulation among the uses for GWM Worlds 2.

Creators and interactive studios are the third, which is where Runway's interactive work and PixVerse's game engine sit. And infrastructure buyers are the fourth: Amazon's write-up ties the growth of world models to a growing need for hardware and infrastructure that can run them, and the AWS and Reactor teams describe their techniques as generalizing across models.

For a creator this week, most of these systems are not yet for sale. Runway Characters, Odyssey-2 Pro and fal's Worlds API can be called today, Genie 3 and GWM Worlds 2 are research previews, and the open weights are a research artefact that runs on a 48GB card. What is worth watching is the direction of the numbers: real-time video above 16 frames per second on a single Trainium2 chip, and a 5B model streaming 832x480 video at 24 fps on one 48GB GPU.

The world models diary

Dated entries, newest first. Each entry is added once and never rewritten; corrections are made in a new entry.

9 Nov 2026

NVIDIA published four open-weight checkpoints for its Cosmos 3 world-model family on 5 November, each a post-trained example of Cosmos 3 Nano with 16B trainable parameters under the OpenMDW-1.1 license. They are simulation tools rather than video generators: Cosmos3-Nano-Transfer-Auto is a bidirectional model that denoises a whole multi-view clip at once for offline synthetic driving data, across up to 11 camera views including fisheyes, with joint RGB and LiDAR generation and up to 400 frames at 480p or 720p; Cosmos3-Nano-Sim-Auto is its causally distilled, few-step streaming variant, returning camera and LiDAR observations to a driving policy in a rolling loop that NVIDIA says also needs a simulation runtime and a World Scenario Map renderer to close it. The other two cover robotics, including a bimanual model that predicts the visual consequences of actions. Sources: huggingface.co, developer.nvidia.com.

Adobe Research and the University of Texas at Austin posted AgiWorld on 2 November, an agentic video world model that treats persistence as the unsolved problem. An agent plans the world and a route through it as a world state and directs the exploration, while a camera-conditioned autoregressive model generates the video five seconds at a time and stores it with its geometry in a point-cloud spatial memory that the agent reads back to update the plan. The authors report state-of-the-art overall results on world-revisiting and exploration benchmarks and minute-long journeys from a single image and a free-form instruction, and argue that persistence comes from the closed loop between the agent and the generator rather than from the video model alone. Sources: bshfang.github.io, arxiv.org.

An unofficial MLX port of NVIDIA's Cosmos 3 Edge video generator appeared on Hugging Face on 7 November, published by NetProjects with a statement that it is not affiliated with NVIDIA. The 8-bit build is 5.0GB with the transformer's linear layers quantized and the VAE kept in BF16, against 7.6GB for the BF16 build, and the publisher's measurement on a MacBook Pro M5 Max with 128GB of memory runs NVIDIA's full image-to-video example at 832x480, 121 frames, 24 fps and 20 steps in 113 seconds with an 18.8GB peak. The port covers the video generator only, and NVIDIA's own card still lists Linux as the sole tested operating system. Source: huggingface.co.

2 Nov 2026

Google DeepMind published Genie 4 on 27 October as a research preview, a world model that generates interactive, photorealistic environments and now generates audio alongside the visuals rather than adding it afterward, so leaves crunch, crowds cheer and characters speak with the player and with each other. The scene can be steered while it runs - describe landscapes and events, move the camera, add characters or change the weather by prompt - and DeepMind's own page carries the caveat that recordings are of real-time interactions with results for illustrative purposes, access is limited to trusted testers, and wider access is promised to Google AI Ultra subscribers. Source: deepmind.google.

A paper submitted to arXiv on 29 October describes L3G-World, a system from KAIST AI, Ewha Womans University and the KAIST Urban AI Institute that generates an explorable world at 9.02 frames per second on a single H100 using a latent 3D Gaussian memory, a set of 3D Gaussians storing VAE latents that renders straight into the generator's latent space instead of being re-encoded from pixels. Its own table reports the best PSNR, LPIPS, FID and subject and background consistency among the world models it tested on the DL3DV benchmark, with memory that grows with the extent of the scene rather than the length of the video; every figure is the authors' measurement of their own system. Source: arxiv.org.

SpAItial's Echo-3, released in October, reconstructs a real space from phone captures as a physically grounded 3D world at real-world scale, keeping what was captured, filling in what was missed and returning clean geometry, per the company's own site. It follows Echo-2, which generated explorable 3D worlds from text or an image, and the company's developer API, which serves persistent 3D Gaussian splat worlds; no independent measurement of Echo-3 has been published. Source: spaitial.ai.

26 Oct 2026

Tencent's Hunyuan team released Hy-Nomas, an open-weight interactive world model, on 23 October: 8 billion parameters, 720p output, sessions up to 60 seconds, multi-turn text control that interleaves with camera moves, and consistent world memory that returns to a previously seen place with its layout intact. The weights and code are Apache 2.0, a suite of four benchmarks with 280 test cases is published alongside, and there is a configuration for running it on a single RTX 4090 with INT8 weights; the project site shows 83 generated worlds across real footage, CG, video-game and animation styles; the technical report is still listed as coming soon and no benchmark scores have been published, so the capability claims are Tencent's own demos. Sources: hy-nomas.hy.tencent.com, huggingface.co, sloptvnews.com.

VAST AI Research open-sourced EdenLive on 18 October, a real-time generative renderer that turns a reference stream - a game-engine graybox, a normal-map animation or recorded footage - plus a target first frame and a prompt into video at 720p and 24 frames per second or better on a single GPU; the 5B model runs on one RTX 5090. It published 5B and 14B weights, bidirectional and few-step chunkwise autoregressive configurations, streaming inference code and EdenGolem v0.1, an agent skill that turns a prompt into an interactive graybox world, and the code is Apache 2.0 with both 5B and 14B weights published on Hugging Face. Sources: tripo3d.ai, huggingface.co, sloptvnews.com.

19 Oct 2026

On 12 October a team from the University of Hong Kong and Tencent's ARC Lab posted Multics to arXiv, a multi-player world model that structures communication across player, group and global levels instead of the fully-connected attention used by Solaris, cutting the quadratic cost of adding players and limiting interference between unrelated events. The paper reports state-of-the-art results at four and six players and says a model trained with four players generalises to six-, eight- and ten-player inference without retraining. Sources: arxiv.org, sloptvnews.com.

On 15 October fal launched Worlds, a browser gallery and API where world models keep generating while a viewer walks through and re-prompts, built on the World Model Accelerator it sells to world-model builders: in-house diffusion-transformer kernels, serverless scaling from one to a thousand GPUs, a WebRTC gateway fal quotes at a median 36 ms frame latency, and a model gallery for distribution. The gallery carries Robbyant's Lingbot World 2 and Alibaba's Abot World 0 as walkable open-weight models and MiniMax's H3 Max Director and H3 Max Reference Director as directable streams; SLOP TV put an hour of generated video at about $4.50. Sources: fal.ai, sloptvnews.com.

12 Oct 2026

On 30 September 2026 Runway announced Praxis-1, its first open-weight world action model, built on the same large-scale video pretraining behind its general world models and aimed at robot control; Runway says it is testing it with early partners and will release it publicly in the coming months, and that simulating policies inside its world model predicts real-world results with 0.95 correlation. Source: runway.com.

On 28 September 2026 AMD said it will acquire World Labs, the spatial-intelligence lab led by Fei-Fei Li, in an all-stock transaction valued at approximately $8.2 billion, expected to close by the end of 2026 subject to regulatory approval. World Labs keeps working on the same research after the close, with Fei-Fei Li joining AMD as executive vice president and chief scientist. Sources: newsroom.amd.com, worldlabs.ai.

26 Sep 2026

The hub opens. This first entry records the news the page starts from, all of it from the labs and the papers rather than from coverage of them.

On 22 September PixVerse announced R2, making input persist so that an action shapes a later moment in the same session, and said its Game Engine now runs on the model. Source: PixVerse.

On 17 September Astronex-World 1.0 was submitted to arXiv: a 5B controllable world model built on the Wan2.2-TI2V-5B prior that generates 832x480 at 24 fps, streams in real time on one Nvidia L20 48GB GPU, and scores 73.5 on WBench Navi and 70.0 on WBench Full. Source: arXiv.

Runway's engineering page for Characters, published on 4 May 2026, describes a real-time video agent built on its GWM-1 general world model that animates a single reference image at 24 fps with 37 ms of model time per frame and 1.75 seconds of server-side turnaround. Source: Runway engineering.

On 3 September Runway showed GWM Worlds 2 as a research preview, generating continuous 720p at 24 fps with 48,000 Hz audio and no preset session length. Source: Runway research.

On 5 August 2025 Google DeepMind announced Genie 3, its first world model to allow real-time interaction, at 720p and 24 frames per second with consistency for a few minutes, and listed its limits as a small action space, weak multi-agent interaction and short sessions. Source: DeepMind.

What changed this week

Against the 2 Nov 2026 edition: NVIDIA's four open Cosmos 3 Nano simulation checkpoints, Adobe Research and UT Austin's AgiWorld paper and the unofficial MLX port of Cosmos 3 Edge for Apple silicon are the week's three additions, all of them physical-AI or research work rather than interactive products for creators. Genie 4, GWM Worlds 2, PixVerse R2 and the hardware facts in the key-facts block were not re-read this week and stand as the 26 September and 2 November record, and neither NVIDIA's checkpoints nor AgiWorld carries an independent benchmark.

SLOP TV's world models coverage

• NVIDIA's 4B Cosmos 3 Edge video model runs on a Mac (9 Nov 2026)
• NVIDIA posts four open Cosmos 3 Nano world models (8 Nov 2026)
• L3G-World generates 3D worlds at 9.02 FPS on one H100 (30 Oct 2026)
• SpAItial's Echo-3 turns phone scans into metric 3D worlds (28 Oct 2026)
• Google's Genie 4 makes photoreal worlds you can ride and hear (27 Oct 2026)
• EdenLive renders 720p video at 24.7 FPS on one RTX 5090 (25 Oct 2026)
• Tencent's Hy-Nomas world model ships with no published scores (24 Oct 2026)
• Matrix-Game 4.0 renders game worlds at 60 FPS on one RTX 5090 (22 Oct 2026)
• SolarRenderer turns game-engine block-outs into video in 4 steps (21 Oct 2026)
• NVIDIA Cosmos3-H-Dreams generates surgical video at 43.2 fps (21 Oct 2026)
• Viggle built a 3D village in one second, and I am sick with envy (19 Oct 2026)
• Amazon in talks to buy AI video startup Decart for $7B (19 Oct 2026)
• fal Worlds streams walkable AI worlds for $4.50 an hour (18 Oct 2026)
• ByteDance's world model re-shoots a movie scene, and Krea 3 lands (15 Oct 2026)
• Helix World generates video and spatial audio in one pass (15 Oct 2026)
• Tencent opens Multics, a world model for 10 players (15 Oct 2026)
• MBZUAI's WorldGuide makes AI video plan its own next steps (12 Oct 2026)
• Seedance 2.5 tops an AI video physics test at 57.76 out of 100 (12 Oct 2026)
• WorldPlay2 turns one image into a playable browser world (8 Oct 2026)
• Runway's Praxis-1 turns video pretraining into robot control (8 Oct 2026)
• Project Continuum guide: Runway's AI operating system explained (1 Oct 2026)
• World-model startup General Intuition raises $220M at $6.2B (30 Sep 2026)
• AMD buys Fei-Fei Li's World Labs for $8.2 billion (30 Sep 2026)
• PixVerse R2 runs a generated world that remembers your moves (25 Sep 2026)
• AWS and Reactor run real-time AI video on Trainium chips (25 Sep 2026)
• Meta Connect 2026: every announcement and its AI video impact (24 Sep 2026)
• fal.live comes back with AI-generated prompts viewers can upvote (23 Sep 2026)
• Runway Labs demos video that redesigns itself to fit the frame (22 Sep 2026)
• World Labs' Atlas steers video from a single photo (22 Sep 2026)
• Meshy's Mora builds a world in 3D before rendering it (22 Sep 2026)
• 5 Real-Time AI Video Systems Ranked, From Shipped API to Research Preview (21 Sep 2026)
• Runway Wants AI Video to Stream as You Prompt It, Frame by Frame (21 Sep 2026)
• Astronex-World 1.0 Streams Open World-Model Video on a Single 48GB GPU (21 Sep 2026)
• Report: ByteDance Is Building a Real-Time Spatial Video Model on Seedance, Targeting 20 fps (19 Sep 2026)
• Runway Reaches $200 Million in Annual Recurring Revenue, Doubling Since April, & Adds Kinetix (19 Sep 2026)
• Victor Riparbelli on Talking Avatars, AI Role-Play & What's Next for Synthesia (Ground Level AI) (19 Sep 2026)