Back to the trend of the day: Google's Genie 4 AI builds worlds to walk through; invited testers only
Poly Rating · from the trend of the day

Best AI World Models 2026: Outside Long-Walk Tests

Genie 3 scores best of all tested models on memory in outside tests but only 8th at following your keys, and Genie 4, announced on October 27, has not been run by a single outside benchmark.

  • No. 1: Genie 3 (Google DeepMind)
  • 10 places

01Final list

The leading AIs each built their own list of ten independently. Here is one result that is easy to check: every place has a reason, a source, and the list of AIs that included it.

Order: from most to least AI agreement.

A bright badge means this AI put the entry on its list. A dim badge means it did not.

  1. 1

    Genie 3 (Google DeepMind)

    #1 first-person on WorldRoamBench (73.81 in the paper, 70.32 on the project site) with the best memory, yet 8th at following keys. Project Genie caps walks at 60 seconds.

    Included it: Claude, ChatGPT, Grok, Nova, Kimi, Qwen, StepFun, DeepSeek. Did not include: Gemini, Wenxin.

    In their lists: Claude: 1 · ChatGPT: 1 · Grok: 1 · Nova: 10 · Kimi: 1 · Qwen: 8 · StepFun: 2 · DeepSeek: 1

    What each AI noted 8
    • Claudeplace 1

      First in memory and physics on WorldRoamBench's first-person board, yet only 8th in action following and 9th in visual quality: the rare model that remembers a room but obeys you less.

      Source ↗
    • ChatGPTplace 1

      The independent arena leader at 1,727 Elo, yet even the leading generation of interactive video worlds remains far from a verified hour-long simulation.

    • Grokplace 1

      Leads outsider long-walk benches (~70-74 FPV overall) while outsiders still see physics/memory gaps versus curated demos.

      Source ↗
    • Novaplace 10

      Foundation for Genie 4, 1-min stability

      Source ↗
    • Kimiplace 1

      The only model that tops both independent walking tests: #1 first-person overall on WorldRoamBench (73.81) and best world-consistency on WorldMark — but your walk ends at 60 seconds by design.

    • Qwenplace 8

      The 2025 predecessor to Genie 4 that pioneered user-steered environments, now largely eclipsed by Cosmos in physics and Marble in 3D export.

      Source ↗
    • StepFunplace 2

      The only world model every outside benchmark runs — and the one whose own one-minute memory claim the tables roughly confirm.

      Source ↗
    • DeepSeekplace 1

      The only world model with a top-three finish on every outside benchmark — and an 8th-place action score that shows exactly where the curated demo ends.

      Source ↗
    8 AIs put it here
  2. 2

    Genie 4 (Google DeepMind)

    Sound, talking characters and typed changes mid-world, but only invited testers can use it and no outside test exists. '2 minutes' comes from press reports; Google gives no number.

    Included it: Claude, ChatGPT, Grok, Nova, Kimi, Qwen, StepFun, DeepSeek. Did not include: Gemini, Wenxin.

    In their lists: Claude: 3 · ChatGPT: 3 · Grok: 3 · Nova: 1 · Kimi: 9 · Qwen: 2 · StepFun: 3 · DeepSeek: 2

    What each AI noted 8
    • Claudeplace 3

      The richest feature list (sound, talking characters, typed changes mid-world) and the thinnest outside evidence: announced yesterday, nobody has benchmarked it.

      Source ↗
    • ChatGPTplace 3

      Roughly two minutes of interactive simulation now includes generated voices and sound, but independent testers have not yet established whether the world survives a return visit.

      Source ↗
    • Grokplace 3

      First DeepMind world model with native audio and dialogue, but still invited-testers-only and unbenchmarked by outsiders.

      Source ↗
    • Novaplace 1

      Native audio, 2-min stability, Gemini-guided planning

      Source ↗
    • Kimiplace 9

      Announced yesterday with the biggest claimed leap — sound, talking characters, a Gemini brain steering the scene — and not a single outside test or published spec to verify it.

      Source ↗
    • Qwenplace 2

      Announced October 27, 2026, it adds real-time native audio and conversational characters, but its true stability remains locked behind an invite-only tester wall.

      Source ↗
    • StepFunplace 3

      The most interactive world model ever shown has not yet been run by anyone outside Google.

      Source ↗
    • DeepSeekplace 2

      Two minutes of coherence, native sound and characters that talk back — and zero outside benchmark rows to prove it.

      Source ↗
    8 AIs put it here
  3. 3

    Marble (World Labs)

    Doesn't drift because the world is a saved 3D file (splats and meshes), not live video. No long-walk benchmark tests it, and it is not a live simulator.

    Included it: Claude, ChatGPT, Gemini, Grok, Kimi, Qwen, StepFun, DeepSeek. Did not include: Nova, Wenxin.

    In their lists: Claude: 6 · ChatGPT: 2 · Gemini: 2 · Grok: 5 · Kimi: 4 · Qwen: 3 · StepFun: 1 · DeepSeek: 3

    What each AI noted 8
    • Claudeplace 6

      It does not stream video: it hands you a persistent Gaussian-splat and mesh world, so turning around is safe by design, but you cannot play inside it for long.

      Source ↗
    • ChatGPTplace 2

      A world you can export as actual 3D geometry has a persistence advantage that beautiful but continuously regenerated video cannot automatically match.

    • Geminiplace 2

      Unlike pure video illusionists, it compiles single photos into exportable 3D Gaussian splat worlds that human architects can walk through indefinitely without visual decay.

    • Grokplace 5

      The one that exports real editable 3D (splats, meshes, GLB) instead of video-only walks.

    • Kimiplace 4

      The cheat code to stability: its worlds never decay because they are exported 3D files — the only top model no benchmark can make 'forget' a room.

    • Qwenplace 3

      The only model that actually exports editable 3D geometry instead of just a video, giving filmmakers and architects a real spatial engine rather than a hallucination.

    • StepFunplace 1

      The only world model whose object permanence is a file format, not a hope.

    • DeepSeekplace 3

      The only top model that exports a Gaussian-splat scene you can reopen in Unreal or Blender — object permanence by geometry, not by memory.

    8 AIs put it here
  4. 4

    NVIDIA Cosmos 3

    Built to train robots and cars and absent from walk tests. NVIDIA's own report says one driving variant holds its metrics for 1 hour (lab claim, two scenes).

    Included it: Claude, ChatGPT, Gemini, Grok, Kimi, Wenxin, Qwen, StepFun. Did not include: Nova, DeepSeek.

    In their lists: Claude: 8 · ChatGPT: 9 · Gemini: 3 · Grok: 4 · Kimi: 8 · Wenxin: 1 · Qwen: 1 · StepFun: 6

    What each AI noted 8
    • Claudeplace 8

      Built to generate training data for robots and cars, not for walking around in, and it is open for developers.

      Source ↗
    • ChatGPTplace 9

      The most consequential world-model platform may be the one building robot-training data, rather than the one winning a beautiful-world video contest.

    • Geminiplace 3

      Built specifically for physical AI rather than gaming, scoring highest on gravity and friction benchmarks while ingesting industrial robot trajectories.

      Source ↗
    • Grokplace 4

      Open omnimodal physical-AI backbone already used for robot/AV sim, ranking high on open physical benches while interactive long-walk scores lag closed models.

      Source ↗
    • Kimiplace 8

      Nobody has walked it for fun, but it is the only model with a published one-hour stability curve — metrics stay in range over a full 60 minutes of driving simulation.

      Source ↗
    • Wenxinplace 1

      The only world model with a published physics engine running robot sims today — minutes of stable worlds for developers.

      Source ↗
    • Qwenplace 1

      The only open model that ranks #1 across seven physical AI leaderboards, turning world generation into a robot-training standard instead of a parlor trick.

    • StepFunplace 6

      Open and omnimodal, but built to train robots rather than to be walked through for an hour.

      Source ↗
    8 AIs put it here
  5. 5

    HY-World 1.5 / WorldPlay (Tencent Hunyuan)

    Follows keys best of all models (98.33 partial accuracy, 70.29 first-person), but third-person control drops to 21.23, against 84.43 for Happy Oyster.

    Included it: Claude, ChatGPT, Grok, Nova, Kimi, Qwen, StepFun, DeepSeek. Did not include: Gemini, Wenxin.

    In their lists: Claude: 4 · ChatGPT: 6 · Grok: 7 · Nova: 2 · Kimi: 6 · Qwen: 6 · StepFun: 5 · DeepSeek: 8

    What each AI noted 8
    • Claudeplace 4

      The only open model sold on 24 FPS plus long-term geometric memory, with full training code and weights, and it was run in WorldMark by outsiders.

      Source ↗
    • ChatGPTplace 6

      An openly released world model with an explicit memory-reconstruction mechanism: the engineering is inspectable even when the generated world is imperfect.

      Source ↗
    • Grokplace 7

      Open weights that lead some action scores yet lag badly on TPV control and physics versus closed models.

      Source ↗
    • Novaplace 2

      Balanced quality and control, open-source

      Source ↗
    • Kimiplace 6

      Follows keyboard commands better than Genie 3 (98.33 partial action accuracy) yet loses the room when a character appears — 21% third-person control rate.

      Source ↗
    • Qwenplace 6

      Tencent’s open-source multi-modal framework builds actual navigable 3D scenes with depth and camera pose, making it a serious rival to closed-source giants.

      Source ↗
    • StepFunplace 5

      The only open model that promises geometric consistency, not just visual memory.

      Source ↗
    • DeepSeekplace 8

      Best navigation in WBench at 87.5, but perspective consistency collapses to 62.5 — a world you can steer but not trust when you turn around.

      Source ↗
    8 AIs put it here
  6. 6

    Matrix-Game 3.0 (Skywork AI)

    The fastest open model (40 FPS at 720p, the authors' claim) scores only 63.31 first-person on outside walks. Its minute-long memory is the lab's own claim.

    Included it: Claude, ChatGPT, Grok, Nova, Kimi, Qwen, StepFun, DeepSeek. Did not include: Gemini, Wenxin.

    In their lists: Claude: 5 · ChatGPT: 7 · Grok: 8 · Nova: 3 · Kimi: 5 · Qwen: 5 · StepFun: 4 · DeepSeek: 7

    What each AI noted 8
    • Claudeplace 5

      720p at up to 40 FPS from a 5B open model, a big jump over 2.0's 25 FPS, but the minute-long claim is the authors' own.

      Source ↗
    • ChatGPTplace 7

      Its 1,487 Elo looks modest beside Genie 3, yet WorldMark identifies it as a strong example of obeying a reverse-direction command.

    • Grokplace 8

      Fast open interactive video model (~5 MPPS) that still trails on long-horizon physics and memory.

      Source ↗
    • Novaplace 3

      Strong physics, third-person support

      Source ↗
    • Kimiplace 5

      The open-source minute-holder: 720p at 40 fps with minute-long memory — and on Oct 15, 2026 it grew into Matrix-Game 4.0, actual playable game levels.

      Source ↗
    • Qwenplace 5

      The open-source champion for long-form 720p video generation, holding scenes together for minutes on WorldMark where proprietary models often dissolve into visual drift.

      Source ↗
    • StepFunplace 4

      About 25 GB of open weights that reaches 40 FPS at 720p on a consumer card.

      Source ↗
    • DeepSeekplace 7

      40 FPS at 720p with a 5B model — the open-source real-time benchmark to beat.

      Source ↗
    8 AIs put it here
  7. 7

    YUME 1.5

    The prettiest frames in WorldMark (imaging 74.36) but weak coherence: 56.21 first-person on WorldRoamBench, 9th of 10. Rooms stop making sense when you turn.

    Included it: Claude, ChatGPT, Grok, Nova, Kimi, Qwen, StepFun, DeepSeek. Did not include: Gemini, Wenxin.

    In their lists: Claude: 9 · ChatGPT: 5 · Grok: 9 · Nova: 4 · Kimi: 10 · Qwen: 7 · StepFun: 10 · DeepSeek: 9

    What each AI noted 8
    • Claudeplace 9

      An open model that WorldMark ran on identical scenes, but with no standout result in the sources I could retrieve.

      Source ↗
    • ChatGPTplace 5

      Third in human preference voting, but WorldMark's action probes show why a convincing scene can still respond poorly to basic movement commands.

    • Grokplace 9

      Strong aesthetic scores on WorldMark but ranks near the bottom on long-horizon WorldRoamBench stability.

      Source ↗
    • Novaplace 4

      High visual appeal, weaker consistency

      Source ↗
    • Kimiplace 10

      The beauty-pageant winner that flunks the memory test: best-looking frames in WorldMark, but rooms stop making sense when you turn around.

      Source ↗
    • Qwenplace 7

      A text-controlled interactive world generator that performs reliably on WorldMark’s unified scenes, though it lacks the real-time audio depth of Genie 4.

      Source ↗
    • StepFunplace 10

      Open and keyboard-playable, and the clearest case of a demo that outruns its own rotations.

      Source ↗
    • DeepSeekplace 9

      The visual-quality winner of WorldMark that ranks poorly on long-horizon coherence — beauty without memory.

      Source ↗
    8 AIs put it here
  8. 8

    Happy Oyster (Alibaba)

    Beats Genie 3 once a character is on screen: #1 third-person on WorldRoamBench (60.24 vs 57.04) and top three in first person. Closed, early access via waitlist.

    Included it: Claude, Grok, Nova, Kimi, Qwen, StepFun, DeepSeek. Did not include: ChatGPT, Gemini, Wenxin.

    In their lists: Claude: 2 · Grok: 2 · Nova: 9 · Kimi: 2 · Qwen: 4 · StepFun: 8 · DeepSeek: 4

    What each AI noted 7
    • Claudeplace 2

      Leads WorldRoamBench's third-person board and sits in the first-person top three, while press says its roaming mode runs about 1 minute.

      Source ↗
    • Grokplace 2

      Tops TPV control and memory on the same outsider walks that expose open-model gaps, with directing mode to 3 minutes.

      Source ↗
    • Novaplace 9

      Early-stage, limited testing

      Source ↗
    • Kimiplace 2

      Beats Genie 3 the moment a character is on screen: #1 third-person on WorldRoamBench with an 84.4% third-person control rate — roughly 4× the best open model.

      Source ↗
    • Qwenplace 4

      A real-time open-ended engine that handles voice and keyboard inputs, scoring high on WorldRoamBench but still struggling to keep physics stable past the 60-second mark.

      Source ↗
    • StepFunplace 8

      A headline name in the most-cited open table — and still behind open models at following your keys.

      Source ↗
    • DeepSeekplace 4

      WorldRoamBench’s third-person champion and WBench’s navigation outlier at 94.8 — then last in visual score.

      Source ↗
    7 AIs put it here
  9. 9

    Lyra 2.0 (NVIDIA)

    The best open model in outside long-walk tests: 70.32 first-person (69.61 on the site), close behind Genie 3. The weights carry a non-commercial licence.

    Included it: Claude, ChatGPT, Grok, Nova, Kimi, DeepSeek. Did not include: Gemini, Wenxin, Qwen, StepFun.

    In their lists: Claude: 7 · ChatGPT: see list · Grok: 6 · Nova: 6 · Kimi: 3 · DeepSeek: 5

    What each AI noted 6
    • Claudeplace 7

      A first-person top-three finisher on WorldRoamBench that gets far less attention than Genie.

      Source ↗
    • ChatGPT

      Named differently in its own list: open the full list.

      see list
    • Grokplace 6

      Highest open-model FPV score on WorldRoamBench (~69-70), close behind closed leaders.

      Source ↗
    • Novaplace 6

      Strong memory, moderate physics

      Source ↗
    • Kimiplace 3

      The best open-weight model in the outsiders' walk test — 70.32 first-person overall, within 3.5 points of Genie 3, on hardware you can rent.

      Source ↗
    • DeepSeekplace 5

      The open-weight leader: second overall in WorldRoamBench and the only top-five model you can download and run.

      Source ↗
    6 AIs put it here
  10. 10

    LingBot-World (Ant Group)

    Steady under hard key sequences (strict accuracy drops under 5 points) and 64.25 first-person on WorldRoamBench, but only 32.19 in third person.

    Included it: ChatGPT, Grok, Nova, Kimi. Did not include: Claude, Gemini, Wenxin, Qwen, StepFun, DeepSeek.

    In their lists: ChatGPT: 8 · Grok: 10 · Nova: 5 · Kimi: 7

    What each AI noted 4
    • ChatGPTplace 8

      A 1,605-Elo contender with 228 arena battles, making its public comparison record larger than several better-known competitors.

    • Grokplace 10

      Open model competitive on action following yet weak on TPV control and overall stability when outsiders run the same walks.

      Source ↗
    • Novaplace 5

      Good spatial consistency, control issues

      Source ↗
    • Kimiplace 7

      The consistent runner-up of the open world: 64.25 on the outsiders' walk test and second only to NVIDIA's own simulator in NVIDIA's camera-navigation benchmark.

      Source ↗
    4 AIs put it here

More about the top three ↓

02More about the top three

The ranking is above. Here is more about the three places at the top: what they are, why they are here, and what is worth knowing.

Only Genie 3 has an outside score (WorldRoamBench). Genie 4 has no outside test, and Marble is a saved 3D file that no long-walk test covers, so the three are not measured the same way.

1Genie 3 (Google DeepMind)

No. 1 — Genie 3 (Google DeepMind). On the lists of 8 of 10 AIs; median place there: 1. It has the best memory of any model in outside walking tests, though it follows your key presses poorly and walks end at 60 seconds.

Did you know

  • Genie 3 can build street scenes starting from Google Street View. Waymo made its own version, the Waymo World Model, to simulate rare road situations for robotaxis. That version outputs lidar four times faster than Genie 3.

What it is
Google's AI that draws a world you walk through live, as video.
Its edge
The only one of the three with outside test scores, first for memory.
Who it suits
Suits people who want to try a live walkable world; not those needing long sessions.
  • WorldRoamBench first-person score73.81 first-person (paper)

Wikipedia · Wikidata

Why it is here
  • In the WorldRoamBench test of 12 world models, Genie 3 came 1st for memory in first-person walks, and 1st for physics too. Its weak spot was following key presses, where it came 8th.

  • Genie 3's memory lasts about one minute, up from 10–20 seconds for Genie 2. But Project Genie, the version people can use, ends each walk at 60 seconds because longer sessions cost too much to offer widely.

    Source ↗ ✓ confirmed by several AIs
Did you know
  • Genie 3 is reported to have 11 billion parameters. Its visual memory of about one minute is three to six times longer than Genie 2's 10–20 seconds. Genie 2 also drew its worlds at only 360p.

History
  • Project Genie opened on January 29, 2026, to Google AI Ultra subscribers in the U.S. Google's page says walks were capped at 60 seconds, and controls could be slow.

Keep in mind
  • TechCrunch says each Project Genie session is capped at one minute because of compute costs. Going back to a place usually worked, but objects could be missing or added when the scene was redrawn.

    Source ↗ ✓ confirmed by several AIs
  • The same review found keys sometimes did nothing or sent the character the wrong way, and characters occasionally walked through walls. Photoreal prompts tended to look like video games.

Where to go further
  • Project Genie runs Genie 3 for US Google AI Ultra subscribers aged 18+. Start with world sketching: describe a scene, pick walking, flying or driving, then move through it. Watch for the 60-second limit.

Compare the top three ↓

2Genie 4 (Google DeepMind)

No. 2 — Genie 4 (Google DeepMind). On the lists of 8 of 10 AIs; median place there: 3. It adds sound and longer worlds, about two minutes by press reports, but only invited testers can use it and no outside test exists.

Did you know

  • Genie 4 appeared as a research preview on October 27, 2026, only a day before this ranking. Google says it is built on Gemini Omni for lifelike detail. DeepMind also frames it as a safe place to test how robots and agents move around.

What it is
Google's newest world model, announced October 27, 2026.
Its edge
Longer worlds than Genie 3 (press reports) and adds sound and talking characters.
Who it suits
Suits those who can join the tester list; others should wait for wider access.
  • WorldRoamBench first-person scoreNo outside test yet

Wikipedia · Wikidata

Why it is here
  • Genie 4 experiences last about two minutes, double the 60-second sessions of the January 2026 Project Genie. That is why it ranks above Genie 3 on length. This figure comes from press reports, not from Google.

    Source ↗ ✓ confirmed by several AIs
  • Google announced Genie 4 on Oct 27, 2026, but only to a limited group of trusted testers. Ultra subscribers get it "soon" and Gemini API partners later, with no date given. That is why it ranks below Genie 3 for who can use it.

    Source ↗ ✓ confirmed by several AIs
History
  • Genie 4 was opened on October 27, 2026, only to trusted testers, as a research preview. The 'up to two minutes' figure comes from TechCrunch reporting and is double Genie 3's limit.

    Source ↗ ✓ confirmed by several AIs
Keep in mind
  • TechCrunch says Genie 4 can keep a scene consistent for up to two minutes, but details and objects can drift the longer it runs. It is only a research preview for trusted testers.

  • TechCrunch reports DeepMind has not yet worked out how to make the audio accurate or any text inside the world readable.

What people say
  • "anytime you’re kind of in need of rapid ideation." — Diego Rivas, Google DeepMind engineer

Where to go further
  • Genie 4 is a research preview for trusted testers, so you probably can't open it yet. Read the page's notes on sound and on 'a few minutes of continuous interaction', then compare with Genie 3 in Project Genie.

Nova put it first. What they saw in it
  • Nova: Native audio, 2-min stability, Gemini-guided planning

Compare the top three ↓

3Marble (World Labs)

No. 3 — Marble (World Labs). On the lists of 8 of 10 AIs; median place there: 3. Its world is a saved 3D file, so it never drifts, but no long-walk test covers it and it is not a live simulator.

Cararra-Steinbruch retouched
Reiner Flassig · Public domain · Wikimedia Commons

Did you know

  • Marble's free plan can make worlds but cannot export them. Saving the splat files, 360 panoramas and physics collider meshes needs a paid plan. The textured GLB mesh needs Pro, and commercial rights come with Pro, not Standard.

What it is
World Labs' tool that turns text or images into saved 3D worlds.
Its edge
Its worlds cannot drift or forget a room, because they are files.
Who it suits
Suits creators and game makers who want exportable 3D; not those wanting a reacting world.
  • WorldRoamBench first-person scoreNo long-walk test

Wikipedia · Wikidata

Why it is here
  • Marble saves each world as a file instead of drawing it live. The full-quality file holds 2 million splats (tiny 3D dots) in .ply or .spz format, so the world stays the same however long you stay.

    Source ↗ ✓ confirmed by several AIs
  • Marble also offers a lighter 500k-splat version made for real-time playback, plus standard 3D meshes. The world is a file you can open in game tools, not a live simulator you walk through.

Did you know
  • A Marble world can leave Marble. Exports go to Unreal Engine, Unity, Blender and Houdini, and to web apps through Spark for three.js. That is why it stays steady: it is a saved file, not video made live.

    Source ↗ ✓ confirmed by several AIs
History
  • In November 2025 World Labs released Marble, its first commercial product, which turns text prompts into downloadable 3D worlds. The company had already reached a valuation of just over $1 billion.

Keep in mind
  • The AI Indigo review says "The most glaring limitation is the lack of granular control." It also says Marble is not a replacement for a full production pipeline.

  • The same review says that while the worlds can be walked through, "the interactivity is limited." So Marble is closer to a saved place to explore than a living, reacting world.

What people say
  • "It doesn’t replace the 3D artist, but it gives every creator the power to build a world." — AI Indigo Team

Where to go further
  • Marble builds a saved 3D world from a text prompt, images or video. Open the Create section, pick a preset or type a short description, then walk around the result and check that the same room is still there when you turn back.

StepFun put it first. What they saw in it
  • StepFun: The only world model whose object permanence is a file format, not a hope.

Compare the top three ↓

The top three side by side

Genie 3 (Google DeepMind)Genie 4 (Google DeepMind)Marble (World Labs)
In the AIs' lists8 / 108 / 108 / 10
WorldRoamBench first-person score73.81 first-person (paper)No outside test yetNo long-walk test
Put first byClaude, ChatGPT, Grok, Kimi, DeepSeekNovaStepFun

Places with the same number of AIs are ordered by their average place in the AIs' own lists.

03Worth knowing

What the AIs noticed while building the list.

The same Genie 3 that ranks 1st in memory and physics ranks 8th in action following and 9th in visual quality: remembering a room and obeying you are different skills.

Noticed by:
Included it: Claude, StepFun, DeepSeek. Did not include: ChatGPT, Gemini, Grok, Nova, Kimi, Wenxin, Qwen.
Source ↗

WorldMark finds visual quality and world consistency largely unrelated: the best-looking model ranks last in direction accuracy and latency.

Noticed by:
Included it: Claude, Kimi, StepFun. Did not include: ChatGPT, Gemini, Grok, Nova, Wenxin, Qwen, DeepSeek.
Source ↗

No model, closed or open, reliably passes action, visuals, physics and memory at once; even the best gets only moderate scores.

Noticed by:
Included it: Grok, Kimi, StepFun. Did not include: Claude, ChatGPT, Gemini, Nova, Wenxin, Qwen, DeepSeek.
Source ↗

Outside tests stop at 60 seconds, so nobody has yet measured whether any generated world holds for an hour.

Noticed by:
Included it: Claude, ChatGPT. Did not include: Gemini, Grok, Nova, Kimi, Wenxin, Qwen, StepFun, DeepSeek.
Source ↗

04Questions people ask

which is the best ai model in the world

Among AI world models, Genie 3 has the strongest outside record: #1 first-person on WorldRoamBench (73.81) with the best memory. Happy Oyster leads third person (60.24).

Source ↗
what is the best ai model in the world

For walkable worlds, Genie 3 leads outside tests. The newer Genie 4 adds sound and dialogue but only invited testers can use it, and it has no outside benchmark yet.

Source ↗
what is the world's best ai model

It depends on the measure. The two WorldRoamBench versions give Genie 3 73.81 or 70.32, and its key-following ranks only 8th, so no model wins on every dimension.

05How we counted

In plain words, no formulas.

1The question

Top 10 AI World Models: Which Ones Hold a World Together When Outsiders Test Them

2Why this topic

It sets outside long-walk and memory benchmarks against the labs' own claims and against who can use each model today, three things no single benchmark or press release shows together.

3Where it applies

global, AI world models that build interactive 3D or video worlds (Genie, Marble, Cosmos, HY-World, Matrix-Game, YUME, Oasis and others)

4What we looked at

  • How long a world stays stable and consistent while you walk through it: minutes, not seconds
  • Object permanence: whether rooms and objects are still the same when you turn around or come back
  • Interaction depth: sound, physics, characters that respond, changes made with plain words
  • Who can use it today: anyone, subscribers only, developers, or invited testers only
  • Whether outside benchmarks have tested it, or only the lab's own claims exist
  • WorldRoamBench (AMAP CV Lab, Alibaba; arXiv 2606.31672, open leaderboard) — How stable 12 interactive world models stay over 10–60 s keyboard walks: action following, visual drift, interaction physics
  • WorldMark (Alaya Lab; arXiv 2604.21686) — The same 500 scenes and action sequences (20–60 s) run on Genie 3, YUME, HY-World, Matrix-Game, Open-Oasis and others, scored on visual quality, control align
  • World Model Arena (Alaya Lab) — Elo ratings from people voting between two models' outputs, 12 models. This measures which output people prefer, not how many minutes a world lasts. Alaya also enters i

5Who ranked

The leading AIs, each building its own list of ten independently and naming a source for every place: Claude, ChatGPT, Gemini, Grok, Nova, Kimi, Wenxin, Qwen, StepFun, DeepSeek.

6How the places were set

The same entry written differently by different AIs was counted as one. The more AIs put an entry on their list, and the higher they put it, the higher its place. "AI agreement" shows how many of the AIs had it on their own list. Sources are only those the AIs found themselves; nothing was added by hand.

06Each AI's own list

🇺🇸 Claude10 places · 4 notes

The same question, answered by Claude alone. The final list above combines all the AIs.

  1. 1
    Genie 3 (Google DeepMind, 2025)

    First in memory and physics on WorldRoamBench's first-person board, yet only 8th in action following and 9th in visual quality: the rare model that remembers a room but obeys you less.

    Source ↗
  2. 2
    Happy Oyster (Alibaba)

    Leads WorldRoamBench's third-person board and sits in the first-person top three, while press says its roaming mode runs about 1 minute.

    Source ↗
  3. 3
    Genie 4 (Google DeepMind, October 2026)

    The richest feature list (sound, talking characters, typed changes mid-world) and the thinnest outside evidence: announced yesterday, nobody has benchmarked it.

    Source ↗
  4. 4
    HY-World 1.5 / WorldPlay (Tencent Hunyuan)

    The only open model sold on 24 FPS plus long-term geometric memory, with full training code and weights, and it was run in WorldMark by outsiders.

    Source ↗
  5. 5
    Matrix-Game 3.0 (Skywork)

    720p at up to 40 FPS from a 5B open model, a big jump over 2.0's 25 FPS, but the minute-long claim is the authors' own.

    Source ↗
  6. 6
    Marble (World Labs)

    It does not stream video: it hands you a persistent Gaussian-splat and mesh world, so turning around is safe by design, but you cannot play inside it for long.

    Source ↗
  7. 7
    Lyra 2.0 (NVIDIA)

    A first-person top-three finisher on WorldRoamBench that gets far less attention than Genie.

    Source ↗
  8. 8
    NVIDIA Cosmos (Cosmos 3, June 2026)

    Built to generate training data for robots and cars, not for walking around in, and it is open for developers.

    Source ↗
  9. 9
    YUME 1.5

    An open model that WorldMark ran on identical scenes, but with no standout result in the sources I could retrieve.

    Source ↗
  10. 10
    Open-Oasis (Oasis)

    The Minecraft-style pioneer that WorldMark now uses as a baseline, a reminder of how far seconds-long worlds have come.

    Source ↗

What Claude noticed

On WorldRoamBench the same Genie 3 that ranks 1st in memory and physics ranks 8th in action following and 9th in visual quality. Memory and looks are different skills.

WorldMark's abstract says the most beautiful model is last in direction accuracy and latency, and the fastest responders are often the least stable.

Genie 4's headline stability is 2 minutes, double Genie 3. WorldRoamBench tests only 10–60 s, so no public test yet reaches hour-long worlds.

Revisit tests are the weak spot: WorldMemBench finds degradation grows with repeated returns to the same place.

🇺🇸 ChatGPT10 places · 2 notes

The same question, answered by ChatGPT alone. The final list above combines all the AIs.

  1. 1
    Google DeepMind Genie 3

    The independent arena leader at 1,727 Elo, yet even the leading generation of interactive video worlds remains far from a verified hour-long simulation.

  2. 2
    World Labs Marble

    A world you can export as actual 3D geometry has a persistence advantage that beautiful but continuously regenerated video cannot automatically match.

  3. 3
    Google DeepMind Genie 4

    Roughly two minutes of interactive simulation now includes generated voices and sound, but independent testers have not yet established whether the world survives a return visit.

    Source ↗
  4. 4
    AlayaWorld

    Only 20 Elo points behind Genie 3 in the public arena, making it a serious challenger rather than an obscure demonstration.

  5. 5
    YUME 1.5

    Third in human preference voting, but WorldMark's action probes show why a convincing scene can still respond poorly to basic movement commands.

  6. 6
    Tencent HY-World 1.5 / HY-WorldPlay

    An openly released world model with an explicit memory-reconstruction mechanism: the engineering is inspectable even when the generated world is imperfect.

    Source ↗
  7. 7
    Skywork Matrix-Game 3.0

    Its 1,487 Elo looks modest beside Genie 3, yet WorldMark identifies it as a strong example of obeying a reverse-direction command.

  8. 8
    Robbyant LingBotWorld-Fast

    A 1,605-Elo contender with 228 arena battles, making its public comparison record larger than several better-known competitors.

  9. 9
    NVIDIA Cosmos 3

    The most consequential world-model platform may be the one building robot-training data, rather than the one winning a beautiful-world video contest.

  10. 10
    Etched Open-Oasis

    The open alternative earns a place for accessibility and reproducibility, even though its arena win rate is only 1.2%.

What ChatGPT noticed

WorldMark uses 500 standardized cases lasting up to 60 seconds. A 60-second test cannot establish that a generated environment remains stable for an hour.

Genie 4's generated speech and audio represent a meaningful interaction advance, but Google itself warns that objects and visual details can drift during longer sessions.

🇺🇸 Gemini10 places · 1 notes

The same question, answered by Gemini alone. The final list above combines all the AIs.

  1. 1
    Google DeepMind Genie 4 (2026)

    Maintains interactive coherence for over 15 continuous minutes with synchronized physics and conversational NPCs, yet remains locked behind Google's private tester portal.

  2. 2
    World Labs Marble (2026)

    Unlike pure video illusionists, it compiles single photos into exportable 3D Gaussian splat worlds that human architects can walk through indefinitely without visual decay.

  3. 3
    NVIDIA Cosmos (2025/2026)

    Built specifically for physical AI rather than gaming, scoring highest on gravity and friction benchmarks while ingesting industrial robot trajectories.

    Source ↗
  4. 4
    Skywork AI Matrix-Game 4.0 (2026)

    Couples an open-source lightweight logic harness with streaming causal diffusion to run playable action worlds at 28 FPS on consumer GPUs.

    Source ↗
  5. 5
    Decart & Etched Oasis (2024/2026)

    The first fully neural interactive Minecraft simulator that anyone could play live in a browser, though items disintegrate once left behind your virtual back.

  6. 6
    Tencent HY-World 2.0 (2026)

    Converts panoramic video prompts into explorable 3D Gaussian Splats with high architectural precision, but stumbles when objects must physically collide.

    Source ↗
  7. 7
    Google DeepMind Genie 3 (2025)

    Demonstrated that neural networks could synthesize controllable platformers from a single image, though long walks invariably melt into surreal dreamscapes.

  8. 8
    UniWorld / UniSim (2025/2026)

    Built specifically for multi-embodiment simulation, teaching virtual arms to manipulate utensils inside generated kitchen rooms with verified collision boxes.

    Source ↗
  9. 9
    YUME Latent World Simulator (2026)

    Operates entirely in an uncompressed continuous latent space, excelling at fluid transition aesthetics while failing basic permanence when objects pass behind trees.

  10. 10
    Danijar Hafner DreamerV3 (2024/2026)

    The granddaddy of model-based reinforcement learning solves Minecraft diamond mining purely inside its imagination, but outputs pixel-abstracted latent dreams rather than photorealistic vistas.

    Source ↗

What Gemini noticed

The divergence between generative video and true 3D spatial models is widening: models that output native 3D representations (like World Labs Marble) achieve over 95% spatial memory retention, whereas 2D diffusion generators (like Oasis) suffer visual collapse within 45 to 60 seconds.

🇺🇸 Grok10 places · 3 notes

The same question, answered by Grok alone. The final list above combines all the AIs.

  1. 1
    Genie 3 (Google DeepMind, 2025)

    Leads outsider long-walk benches (~70-74 FPV overall) while outsiders still see physics/memory gaps versus curated demos.

    Source ↗
  2. 2
    Happy Oyster (Alibaba ATH, 2026)

    Tops TPV control and memory on the same outsider walks that expose open-model gaps, with directing mode to 3 minutes.

    Source ↗
  3. 3
    Genie 4 (Google DeepMind, October 2026)

    First DeepMind world model with native audio and dialogue, but still invited-testers-only and unbenchmarked by outsiders.

    Source ↗
  4. 4
    NVIDIA Cosmos 3 (2026)

    Open omnimodal physical-AI backbone already used for robot/AV sim, ranking high on open physical benches while interactive long-walk scores lag closed models.

    Source ↗
  5. 5
    Marble (World Labs, 2025-2026)

    The one that exports real editable 3D (splats, meshes, GLB) instead of video-only walks.

  6. 6
    Lyra 2.0 (NVIDIA, 2026)

    Highest open-model FPV score on WorldRoamBench (~69-70), close behind closed leaders.

    Source ↗
  7. 7
    HY-World 1.5 (Tencent Hunyuan, 2025-2026)

    Open weights that lead some action scores yet lag badly on TPV control and physics versus closed models.

    Source ↗
  8. 8
    Matrix-Game 3.0 (Skywork AI, 2026)

    Fast open interactive video model (~5 MPPS) that still trails on long-horizon physics and memory.

    Source ↗
  9. 9
    YUME 1.5 (Shanghai AI Lab, 2025)

    Strong aesthetic scores on WorldMark but ranks near the bottom on long-horizon WorldRoamBench stability.

    Source ↗
  10. 10
    LingBot-World (Ant Group, 2026)

    Open model competitive on action following yet weak on TPV control and overall stability when outsiders run the same walks.

    Source ↗

What Grok noticed

No model dominates every WorldRoamBench dimension: closed models lead physics and memory, several open models match or beat them on per-frame action following.

Genie 4’s ‘two minutes’ figure appears in news coverage but not as a hard number on DeepMind’s own pages, which still say ‘a few minutes’—the same phrasing used for Genie 3.

Marble is the clearest production path for professionals because it exports editable 3D assets rather than video streams that cannot leave the tool.

🇺🇸 Nova10 places · 3 notes

The same question, answered by Nova alone. The final list above combines all the AIs.

  1. 1
    Genie 4 (Google DeepMind)

    Native audio, 2-min stability, Gemini-guided planning

    Source ↗
  2. 2
    HY-World 1.5

    Balanced quality and control, open-source

    Source ↗
  3. 3
    Matrix-Game 3.0

    Strong physics, third-person support

    Source ↗
  4. 4
    YUME 1.5

    High visual appeal, weaker consistency

    Source ↗
  5. 5
    LingBot World

    Good spatial consistency, control issues

    Source ↗
  6. 6
    Lyra 2.0

    Strong memory, moderate physics

    Source ↗
  7. 7
    SANA-WM

    Balanced performance, open-source

    Source ↗
  8. 8
    minWM

    Compact, efficient inference

    Source ↗
  9. 9
    Happy Oyster

    Early-stage, limited testing

    Source ↗
  10. 10
    Genie 3

    Foundation for Genie 4, 1-min stability

    Source ↗

What Nova noticed

Genie 4's native audio and Gemini integration create a new standard for multimodal world models, but remain invite-only.

Open-source models like HY-World and Matrix-Game lead in transparency but lag closed-source systems in long-horizon stability.

WorldRoamBench reveals a trade-off: models optimizing physics often deviate from prescribed trajectories.

🇨🇳 Kimi10 places · 3 notes

The same question, answered by Kimi alone. The final list above combines all the AIs.

  1. 1
    Genie 3 (Google DeepMind, 2025)

    The only model that tops both independent walking tests: #1 first-person overall on WorldRoamBench (73.81) and best world-consistency on WorldMark — but your walk ends at 60 seconds by design.

  2. 2
    Happy Oyster (Alibaba, 2026)

    Beats Genie 3 the moment a character is on screen: #1 third-person on WorldRoamBench with an 84.4% third-person control rate — roughly 4× the best open model.

    Source ↗
  3. 3
    Lyra 2.0 (NVIDIA, 2026)

    The best open-weight model in the outsiders' walk test — 70.32 first-person overall, within 3.5 points of Genie 3, on hardware you can rent.

    Source ↗
  4. 4
    Marble (World Labs, 2025)

    The cheat code to stability: its worlds never decay because they are exported 3D files — the only top model no benchmark can make 'forget' a room.

  5. 5
    Matrix-Game 3.0 (Skywork, 2026)

    The open-source minute-holder: 720p at 40 fps with minute-long memory — and on Oct 15, 2026 it grew into Matrix-Game 4.0, actual playable game levels.

    Source ↗
  6. 6
    HY-World 1.5 / HY-World 2.0 (Tencent, 2025–2026)

    Follows keyboard commands better than Genie 3 (98.33 partial action accuracy) yet loses the room when a character appears — 21% third-person control rate.

    Source ↗
  7. 7
    LingBot-World (Ant Group, 2026)

    The consistent runner-up of the open world: 64.25 on the outsiders' walk test and second only to NVIDIA's own simulator in NVIDIA's camera-navigation benchmark.

    Source ↗
  8. 8
    NVIDIA Cosmos 3 (NVIDIA, 2026)

    Nobody has walked it for fun, but it is the only model with a published one-hour stability curve — metrics stay in range over a full 60 minutes of driving simulation.

    Source ↗
  9. 9
    Genie 4 (Google DeepMind, 2026)

    Announced yesterday with the biggest claimed leap — sound, talking characters, a Gemini brain steering the scene — and not a single outside test or published spec to verify it.

    Source ↗
  10. 10
    YUME 1.5 (2025–2026)

    The beauty-pageant winner that flunks the memory test: best-looking frames in WorldMark, but rooms stop making sense when you turn around.

    Source ↗

What Kimi noticed

Every published walking test agrees nobody passes: WorldRoamBench's authors conclude no model reliably satisfies action, vision, physics and memory simultaneously — 'even the best achieves only moderate scores' across 1,000+ test cases.

Pretty and stable are different skills: WorldMark found visual quality and world consistency 'largely uncorrelated' — YUME makes the most beautiful frames while Genie 3 keeps the most coherent rooms, and precise control (HY-GameCraft) can come at the cost of visual fidelity.

Turning around is the hardest move in the field: NVIDIA's LoopMem benchmark shows sliding-window memory forgets scene structure after a few seconds on revisit, and WorldRoamBench's third-person control rates range from 84% (Happy Oyster) to 21% (HY-World) — the widest gap of any measured dimension.

🇨🇳 Wenxin10 places · 3 notes

The same question, answered by Wenxin alone. The final list above combines all the AIs.

  1. 1
    NVIDIA Cosmos (2026)

    The only world model with a published physics engine running robot sims today — minutes of stable worlds for developers.

    Source ↗
  2. 2
    Google DeepMind Genie 4 (2026)

    Lab says minutes with sound and talking characters — but no outsider has walked it yet; invited-only since Oct 27.

  3. 3
    Google DeepMind Genie 3 (2024–25)

    First 'you can walk through it' demo that outsiders could probe — WorldRoamBench put it at the top of open tests in 2025.

  4. 4
    HY-World (2026)

    Open-source world model that holds 2–3 minute walks in community tests — a rare feat for code you can clone.

    Source ↗
  5. 5
    YUME (2025–26)

    YouTube-backed model that generates video worlds on command — WorldMark tests show good visuals, weaker physics.

    Source ↗
  6. 6
    Marble (Happy Oyster, 2026)

    Community favourite for open-source long walks — GitHub leaderboard shows it edging past HY-World on some scenes.

  7. 7
    Matrix-Game (2026)

    Game-engine-born world model that exports editable 3D — WorldMark scores it well on action consistency.

    Source ↗
  8. 8
    Oasis (2025–26)

    Meta-backed open model that generates whole 3D scenes — less stable over long walks than top tier but free to try.

    Source ↗
  9. 9
    Genie 2 / Genie 3 (legacy, 2024)

    The model that started the 'walk-through' craze — now outscored but still the baseline outsiders compare against.

  10. 10
    WorldBench baseline open models (ensemble, 2026)

    Not one model but the best open-source bundle tested — proof that collective code can reach 40-second walks.

What Wenxin noticed

A 60-second stable world is the current ceiling for open models; Genie 4 claims 'minutes' but no outsider has timed it yet — the gap between lab demo and benchmark reality is the story of 2026.

NVIDIA Cosmos is the only world model with a real physics engine verified in robotics sims — it trades visual flair for physical correctness, which is why it leads outside entertainment.

WorldMark is the first benchmark that forces every lab to run the same 50 scenes — before it, each company picked its own 'best demo' and nobody could compare.

🇨🇳 Qwen10 places · 3 notes

The same question, answered by Qwen alone. The final list above combines all the AIs.

  1. 1
    NVIDIA Cosmos 3

    The only open model that ranks #1 across seven physical AI leaderboards, turning world generation into a robot-training standard instead of a parlor trick.

  2. 2
    Genie 4 (Google DeepMind, 2026)

    Announced October 27, 2026, it adds real-time native audio and conversational characters, but its true stability remains locked behind an invite-only tester wall.

    Source ↗
  3. 3
    World Labs Marble

    The only model that actually exports editable 3D geometry instead of just a video, giving filmmakers and architects a real spatial engine rather than a hallucination.

  4. 4
    Happy Oyster (Alibaba, 2026)

    A real-time open-ended engine that handles voice and keyboard inputs, scoring high on WorldRoamBench but still struggling to keep physics stable past the 60-second mark.

    Source ↗
  5. 5
    Matrix-Game 4.0 (Skywork AI)

    The open-source champion for long-form 720p video generation, holding scenes together for minutes on WorldMark where proprietary models often dissolve into visual drift.

    Source ↗
  6. 6
    HY-World 2.0 (Tencent Hunyuan)

    Tencent’s open-source multi-modal framework builds actual navigable 3D scenes with depth and camera pose, making it a serious rival to closed-source giants.

    Source ↗
  7. 7
    YUME 1.5

    A text-controlled interactive world generator that performs reliably on WorldMark’s unified scenes, though it lacks the real-time audio depth of Genie 4.

    Source ↗
  8. 8
    Genie 3 (Google DeepMind)

    The 2025 predecessor to Genie 4 that pioneered user-steered environments, now largely eclipsed by Cosmos in physics and Marble in 3D export.

    Source ↗
  9. 9
    DIAMOND (Diffusion World Model)

    A reinforcement-learning world model built to master Atari games by simulating them, proving that internal physics engines still beat generative video for pure gameplay logic.

    Source ↗
  10. 10
    Oasis (Decart)

    Famous for generating playable Minecraft-like worlds at 30 frames per second, but independent tests show its object permanence drops sharply after a few room transitions.

What Qwen noticed

A world you can walk through for 60 seconds is a different product from one you can stay in for an hour; WorldRoamBench proves that most models suffer from severe visual drift after just 10 seconds of continuous movement.

Open-source models like Matrix-Game and HY-World are actually beating proprietary labs on standardized tests like WorldMark, because they optimize for memory-augmented geometry rather than just pretty video frames.

The gap between a lab's curated demo and an outside benchmark is massive: Genie 4's launch video shows flawless physics, but independent testers report object permanence fails the moment a character turns around.

🇨🇳 StepFun10 places · 4 notes

The same question, answered by StepFun alone. The final list above combines all the AIs.

  1. 1
    Marble (World Labs)

    The only world model whose object permanence is a file format, not a hope.

  2. 2
    Genie 3 (Google DeepMind)

    The only world model every outside benchmark runs — and the one whose own one-minute memory claim the tables roughly confirm.

    Source ↗
  3. 3
    Genie 4 (Google DeepMind, 2026)

    The most interactive world model ever shown has not yet been run by anyone outside Google.

    Source ↗
  4. 4
    Matrix-Game 3.0 (Skywork AI)

    About 25 GB of open weights that reaches 40 FPS at 720p on a consumer card.

    Source ↗
  5. 5
    HY-World 2.0 / WorldPlay (Tencent Hunyuan)

    The only open model that promises geometric consistency, not just visual memory.

    Source ↗
  6. 6
    NVIDIA Cosmos 3

    Open and omnimodal, but built to train robots rather than to be walked through for an hour.

    Source ↗
  7. 7
    AlayaWorld (Alaya Lab, 2026)

    The benchmark authors' own answer to the table they built.

    Source ↗
  8. 8
    Happy Oyster (Alibaba, 2026)

    A headline name in the most-cited open table — and still behind open models at following your keys.

    Source ↗
  9. 9
    Oasis 3 (Decart)

    Sub-40ms latency buys you reaction time, not an hour of memory.

    Source ↗
  10. 10
    YUME 1.5 (Alaya)

    Open and keyboard-playable, and the clearest case of a demo that outruns its own rotations.

    Source ↗

What StepFun noticed

Memory, not picture quality, is what breaks first: WorldRoamBench's mean overall score falls from 63.17 at short horizon, and memory is the most horizon-sensitive dimension while the mean action score decays only 3.6%.

Pretty is not controllable: WorldMark found perceptual quality correlates positively with all three memory metrics (rho = 0.76-0.79) yet negatively with action metrics, so the best-looking world can be the worst at obeying you.

No model passes everything: WorldRoamBench's authors evaluated 10+ open and closed models and report that none reliably satisfies all four dimensions, with even the best achieving only moderate scores.

The rotation gap is the whole story in one number: across the ten models WorldMark drives, rotational stability spans 31.7 to 87.2 on identical action sequences.

🇨🇳 DeepSeek10 places · 3 notes

The same question, answered by DeepSeek alone. The final list above combines all the AIs.

  1. 1
    Genie 3 (Google DeepMind, 2025–2026)

    The only world model with a top-three finish on every outside benchmark — and an 8th-place action score that shows exactly where the curated demo ends.

    Source ↗
  2. 2
    Genie 4 (Google DeepMind, announced Oct 27, 2026)

    Two minutes of coherence, native sound and characters that talk back — and zero outside benchmark rows to prove it.

    Source ↗
  3. 3
    Marble (World Labs, 2025–2026)

    The only top model that exports a Gaussian-splat scene you can reopen in Unreal or Blender — object permanence by geometry, not by memory.

  4. 4
    Happy Oyster (Alibaba, 2026)

    WorldRoamBench’s third-person champion and WBench’s navigation outlier at 94.8 — then last in visual score.

    Source ↗
  5. 5
    Lyra 2.0 (NVIDIA, 2026)

    The open-weight leader: second overall in WorldRoamBench and the only top-five model you can download and run.

    Source ↗
  6. 6
    Odyssey-3 (Odyssey, 2026)

    Physics-IQ Verified’s highest video-to-video score, 66.1 — but its WorldMark wins are self-run, not outsider-run.

  7. 7
    Matrix-Game 3.0 (Tencent, 2026)

    40 FPS at 720p with a 5B model — the open-source real-time benchmark to beat.

    Source ↗
  8. 8
    HY-World 1.5 (Tencent, 2025–2026)

    Best navigation in WBench at 87.5, but perspective consistency collapses to 62.5 — a world you can steer but not trust when you turn around.

    Source ↗
  9. 9
    YUME 1.5 (2025–2026)

    The visual-quality winner of WorldMark that ranks poorly on long-horizon coherence — beauty without memory.

    Source ↗
  10. 10
    Oasis 3 (Decart, 2026)

    The only world model with a live API at $0.02 per second — built for driving, not walking through rooms.

What DeepSeek noticed

Genie 3 leads WorldRoamBench overall at 70.32, but is only 8th in action score and 9th in visual score — the overall rank hides dimension-level failures that matter for walking and turning.

Happy Oyster’s navigation score in WBench is 94.8 against Genie 3’s 32.2, yet Genie 3 scores 100.0 on causal fidelity — steering accuracy and physical cause-and-effect are almost separate products.

Genie 4’s two-minute coherence is double Genie 3’s, but no outside benchmark has tested it; the gap between the lab demo and the public table is at its widest in this generation.