Not Video Generation, but World Building
Start from an image or prompt and directly generate real 3D worlds (Mesh / 3DGS). Not a short-lived clip, but a browsable, editable, and reusable spatial asset.
Early access — HY-World 2.0 rolling out
HY World Model 2
Start from a single image or prompt and generate a realistic, explorable 3D world that is navigable, editable, and persistent. Powered by HY World Model 2, creation goes beyond visuals and into space.
StatusWorldMirror 2.0 weights are live; HY-Pano 2.0 and WorldStereo 2.0 roll out next. Tencent's official preview is at sceneTo3D playground — free to view, but sign-up currently requires a Mainland China mobile number.
Preview — launching soon
Creation is locked during early access. Join the waitlist to be first in line.
Get notified when global sign-up opens — no China number required.
Prompt-driven scene building
One environmentmultiple shots, one visual logic
Beyond Single Frames, Toward Full Spatial Expansion
With panoramic generation, path planning, and world expansion, AI fills areas beyond your current view to create coherent and navigable full 3D scenes, not isolated fragments.
Not Just Watching, but Free Exploration
Supports first-person and third-person perspectives with free movement and interaction in generated worlds. Built-in physical feedback helps users truly step into AI-created space.
What is HY World Model?
HY World Model (also known as HunyuanWorld) is Tencent's family of open-source AI models that generate immersive, explorable, and interactive 3D worlds from text prompts or single images. "Hunyuan" (混元) means "primordial unity" in Chinese philosophy — later shortened to "HY" for international branding.
Also known as
HunyuanWorld 2.0 · Hunyuan World Model · HY-World 2 · Tencent HY · Hunyuan · 混元
Real 3D Output
HunyuanWorld 2.0 generates actual 3D scenes — not just video. Outputs are directly editable and usable inside game engines like Unreal Engine and Unity.
Fully Open-Source
Model weights, training code, and inference pipelines are all open-sourced on GitHub. Run it on your own hardware or fine-tune it for custom domains.
Top-Ranked Benchmark
The HunyuanWorld-Voyager variant ranks #1 on Stanford's WorldScore benchmark for both video generation and 3D reconstruction quality.
Version Timeline
From the first open-source 3D world model to real-time interactive worlds and native game-engine output.
April 16, 2026
HY-World 2.0 / HunyuanWorld 2.0 — Released
Multi-modal world model for generation, reconstruction, and simulation of 3D worlds. Four-stage pipeline (HY-Pano 2.0 → WorldNav → WorldStereo 2.0 → WorldMirror 2.0) outputs real Mesh / 3DGS / point clouds directly usable in Blender, Unity, Unreal Engine, and Isaac Sim. WorldMirror 2.0 code and weights are open-source under the tencent-hy-world-2.0-community license.
March 8, 2026
HY-World 1.5 — WorldCompass RL Code Released
Open-sourced the reinforcement learning post-training code (WorldCompass) for the WorldPlay-8B model, enabling community fine-tuning.
January 2026
HY-World 1.5 — Training Code Released
Full training code for the WorldPlay-8B model open-sourced, enabling the community to train their own world models from scratch.
December 17, 2025
HY-World 1.5 (WorldPlay)
First open-source real-time interactive world model with long-term geometric consistency. Streams video at 24 FPS with keyboard/mouse control. Built on the HunyuanVideo-8B backbone.
October 22, 2025
HunyuanWorld 1.1 (WorldMirror)
Universal feed-forward 3D reconstruction model. Video-to-3D and multi-view image-to-3D generation. Outputs dense point clouds, depth maps, and 3D Gaussian Splatting.
September 2, 2025
HunyuanWorld-Voyager
First ultra-long-range world model with native 3D reconstruction. Exports point cloud videos to 3D formats without external tools. #1 on Stanford WorldScore.
July 26, 2025
HunyuanWorld 1.0
Industry's first open-source 3D world generation model. Text/image to immersive, explorable 3D worlds. Compatible with CG pipelines. Lite version runs on consumer GPUs (24GB VRAM).
Inside HY-World 2.0
HY-World 2.0 is a multi-modal world model framework. It takes text, single-view images, multi-view images, or video, and outputs real 3D assets — meshes, 3D Gaussian Splatting (3DGS), point clouds, depth maps, surface normals, and camera parameters — in a single pass.
Stage 01
HY-Pano 2.0
Panorama generation. Converts a text prompt or a single image into a 360° panorama that anchors the world's style, lighting, and layout. Supports realistic, cartoon, and game styles.
Stage 02
WorldNav
Spatial planning. A spatial agent predicts camera trajectories from the panorama — what you can walk through, look at, and explore — before any 3D is synthesized.
Stage 03
WorldStereo 2.0
World expansion. Generates a navigable 3DGS world along the planned trajectories with tight camera control — RotErr 0.492, TransErr 0.968, ATE 1.768, outperforming WorldStereo 1.0 on all three.
Stage 04
WorldMirror 2.0 + 3DGS Learning
World composition and reconstruction. A ~1.2B-parameter unified feed-forward model predicts depth, surface normals, camera parameters, point clouds, and 3DGS attributes in a single forward pass — at resolutions from 50K to 500K pixels.
Model Zoo
| Model | Role | Params | Status |
|---|---|---|---|
| WorldMirror 2.0 | Multi-view / video → 3D reconstruction | ~1.2B | Released |
| WorldMirror 1.0 | Legacy multi-view / video → 3D | ~1.2B | Released |
| HY-PanoGen (HY-Pano 2.0) | Text / image → 360° panorama | — | Coming soon |
| WorldStereo 2.0 | Panorama → navigable 3DGS world | — | Coming soon |
WorldMirror 2.0Released
Multi-view / video → 3D reconstruction
~1.2B
WorldMirror 1.0Released
Legacy multi-view / video → 3D
~1.2B
HY-PanoGen (HY-Pano 2.0)Coming soon
Text / image → 360° panorama
—
WorldStereo 2.0Coming soon
Panorama → navigable 3DGS world
—
Benchmarks
WorldMirror 2.0 delivers best-in-class point-cloud reconstruction and WorldStereo 2.0 leads on camera control — the first open-source 3D world model in the same tier as closed-source systems like Marble.
WorldMirror 2.0 — Point-map Reconstruction (high-res + all priors)
| Dataset | Accuracy ↓ | Completeness ↓ |
|---|---|---|
| 7-Scenes | 0.012 | 0.016 |
| NRGBD | 0.015 | 0.016 |
| DTU | 0.554 | 0.771 |
Highlights
- HunyuanWorld-Voyager previously set #1 on Stanford WorldScore at 77.62 (vs. WonderWorld 72.69, CogVideoX-I2V 62.15).
- Cumulative downloads across the Hunyuan 3D series exceeded 3 million on the open-source community by March 2026.
- Integrated into Tencent's internal game pipeline (DreamStar, Light Game Dream Workshop) and adopted by Maxon Cinema 4D.
Technical Specifications
Architecture, parameters, and capabilities of the HY World Model family.
HY-World 2.0 Architecture
- Inputs
- Text · single image · multi-view images · video
- Outputs
- Mesh · 3DGS · point clouds · depth · surface normals · camera params
- Generation pipeline
- HY-Pano 2.0 → WorldNav → WorldStereo 2.0 → WorldMirror 2.0 + 3DGS learning
- Reconstruction core
- WorldMirror 2.0 — unified feed-forward, single forward pass
- WorldMirror 2.0 params
- ~1.2B
- Inference resolution
- 50K – 500K pixels (flexible)
- Styles
- Realistic · cartoon · game (text-driven)
Runtime & Deployment
- License
- tencent-hy-world-2.0-community (open source)
- CUDA
- 12.4 recommended
- Python / PyTorch
- Python 3.10 · PyTorch 2.4.0
- Attention
- FlashAttention-3 (preferred) or FlashAttention-2
- Multi-GPU
- FSDP with BF16 precision
- Target engines
- Blender · Unity · Unreal Engine · Isaac Sim
- Character Mode
- First/third-person exploration with physics-based collision
How It Compares
HY World Model vs. other leading AI world and video generation models.
HY World Model
Tencent
- Open-source (weights + code)
- Real 3D output (game-engine ready)
- Real-time 24 FPS generation
- Interactive keyboard/mouse control
- Consumer GPU support (24 GB)
- #1 Stanford WorldScore
- CG pipeline compatible
Sora
OpenAI
- Closed-source
- Video only (no 3D output)
- Up to 1 minute video
- No interactive control
- Cloud-only
- Physics-accurate motion
- Dynamic camera motion
Veo 3
Google DeepMind
- Closed-source
- Video only (no 3D output)
- Up to 60 seconds
- No interactive control
- Limited public access
- Cinematic quality
- Complex camera transitions
Genie 2
Google DeepMind
- Closed-source
- Video only (no true 3D)
- Interactive playable worlds
- Action-conditional generation
- Not publicly available
- Persistent memory
- Not open-source
Latest News & Updates
Official announcements and community reactions.
Official Announcements
HY
Tencent HY
@TencentHunyuan
HunyuanWorld-Voyager is here and fully open-source! The world's first ultra-long-range world model with native 3D reconstruction, redefining AI-driven spatial intelligence for VR, gaming, and simulations. Direct 3D Output: exports point cloud videos to 3D formats without tools.
HY
Tencent HY
@TencentHunyuan
We're thrilled to release and open-source Hunyuan3D World Model 1.0! This model lets you generate immersive, explorable, and interactive 3D worlds from just a sentence or an image. It's the industry's first open-source 3D world generation model, compatible with CG pipelines.
HY
Tencent HY
@TencentHunyuan
Staying on top! HunyuanWorld-Voyager is now #1 trending on Hugging Face! Our open-source world model is also top-ranked on Stanford's WorldScore, showcasing its strength in both video and 3D reconstruction.
In the News
Tencent Releases Hunyuan World Model 1.5 for Real-Time Interactive Virtual Worlds
AIBase — December 2025
Real-Time AI World Generation: Tencent's HY-World 1.5
AIFilms Studio
HY World 1.5: The Dawn of Text-to-World Simulation
XROM
Tencent Folds AI Lab Into Hunyuan Team in Major AI Overhaul
Caixin Global — March 2026
Tencent to Launch Hunyuan 3.0 in April, Build WeChat AI Agent
Caixin Global — March 2026
Tencent Releases HY-World 1.5 WorldPlay at 24 FPS on Hugging Face
AIGazine
Open-Source Resources
GitHub repositories, model weights, technical reports, and playgrounds.
Repositories
HY-World 2.0 on GitHub
Multi-modal world model — generation, reconstruction, simulation (code + WorldMirror 2.0 weights)
HunyuanWorld-Mirror
Fast and universal 3D reconstruction model (WorldMirror family)
HY-WorldPlay (v1.5)
Real-time interactive world model — code, training, inference
HunyuanWorld 1.0
First open-source 3D world generation model
HunyuanWorld-Voyager
Ultra-long-range world model with native 3D reconstruction
HunyuanVideo
13B/8.3B video generation backbone model
Tencent-Hunyuan GitHub Org
All Hunyuan open-source projects in one place
Models & Resources
HY-World 2.0 on Hugging Face
Official v2 model card — weights + model zoo
HY-World 2.0 Technical Report (PDF)
Four-stage pipeline, architecture, benchmark results
HY-World 2.0 Paper on Hugging Face
Paper discussion and citations
Try HY-World 2.0 — SceneTo3D
Official online playground from Tencent Hunyuan
Try HunyuanWorld on fal.ai
Third-party playground — image-to-world generation
HY-WorldPlay on Hugging Face
v1.5 model weights for local deployment
HunyuanWorld-1 on Hugging Face
v1.0 model weights
HY-World Discord
Community Discord for questions, demos, and issues
@TencentHunyuan on X
Official announcements and demos
How to Use the HY World Model 2 Generator
HY World Model 2 is easy to use in just three simple steps.
Step 01
Define Your World
Choose a preset image or upload your own asset as the starting point and visual foundation for your 3D world. Landscapes, interiors, or conceptual scenes can all serve as your initial world framework.
Step 02
Build and Create
Enter a prompt describing what you want to generate and expand. Based on your intent, AI completes the spatial structure, details, and style to construct a full 3D world.
Step 03
Wait for Generation
In about 3-5 minutes, your explorable 3D world is ready. Freely navigate and switch viewpoints, turning your ideas into a real spatial experience.
An AI World Generator for Everyone
From a single image to a complete world, from raw inspiration to explorable space, HY World Model 2 provides powerful and unified 3D generation capabilities for all kinds of creators, helping every idea become buildable and real.
Filmmakers and Video Creators
Build repeatable and controllable 3D scenes instead of one-off visual outputs. Revisit the same world for multiple shots, precisely adjust camera paths, composition, and spatial relationships, and keep every frame consistent for more efficient storytelling.
Social Media and Virtual IP Creators
Create a dedicated world for your virtual persona. Daily content, story-driven posts, and brand communication can all happen in one persistent 3D space, making your visual style more unified, memorable, and recognizable over time.
Game Designers and Concept Artists
Quickly turn ideas into explorable 3D scenes rather than static concept art. From atmosphere to spatial layout, generate and navigate worlds instantly to validate worldbuilding and level design early in the process.
Architects and Spatial Designers
Transform design ideas directly into navigable spatial prototypes. With a simple image or text description, generate 3D scenes with structure and scale for concept presentations, ideation, and client communication.
F A Q
What is HY World Model 2?
HY World Model 2 is a multimodal 3D world generation framework built on Hunyuan foundation models. It can generate explorable and editable 3D worlds directly from text or a single image, not just video clips.
What kinds of content can HY World Model 2 generate?
HY World Model 2 can generate complete 3D worlds from a single image or prompt, including structured assets such as Mesh and 3D Gaussian Splatting (3DGS). These worlds support free navigation, viewpoint switching, and diverse creative or development workflows.
How is HY World Model 2 different from traditional video generation models?
Unlike traditional video models (such as Genie or HY World 1.5) that generate one-off footage, HY World Model 2 creates persistent 3D worlds. Users can enter, explore, edit, and reuse these spatial assets across different projects.
Can generated 3D worlds be imported into other tools or engines?
Yes. Assets generated by HY World Model 2 (such as Mesh and 3DGS) can be used in mainstream 3D tools and engines, including Unity and Unreal Engine, for further editing and production.
Do I need an account to use HY World Model 2?
Yes. You need to register and sign in to access core features, including 3D world generation and world management.
How is HY World Model 2 priced?
HY World Model 2 uses a usage-based pricing model. Each 3D world generation consumes a fixed cost, and the exact pricing is clearly displayed in the product interface.
How long does it take to generate a 3D world?
In most cases, generating a complete 3D world with HY World Model 2 takes about 3-5 minutes. Actual time may vary depending on scene complexity and content scope.
Will HY World Model 2 keep evolving? What improvements are planned?
Yes. HY World Model 2 will continue to evolve with better world quality, richer scene types, stronger interaction capabilities, and deeper world-building control to make generated 3D worlds more realistic and practical.
What is the difference between HY World, HunyuanWorld, and Hunyuan?
"Hunyuan" (混元, "primordial unity") is Tencent's umbrella AI brand, rebranded to "Tencent HY" for global markets. "HunyuanWorld" or "HY World Model" refers specifically to the world-generation product line. Other Hunyuan models include HunyuanVideo (video generation), Hunyuan3D (3D assets), and the Hunyuan LLM. Common aliases: HY-World, HY World Model, Hunyuan World Model, HunyuanWorld.
Is HY World Model open source? Where can I download it?
Yes, the entire HY World Model family is open-source. Repositories are hosted at github.com/Tencent-Hunyuan. Key repos include HY-WorldPlay (v1.5), HunyuanWorld-1.0, and HunyuanWorld-Voyager. Model weights are available on Hugging Face under the "tencent" organization. HunyuanWorld 2.0 is scheduled to open-source on April 15, 2026.
What hardware do I need to run HY World Model?
The recommended pipeline uses the HunyuanVideo-8B backbone and benefits from GPUs with 40+ GB VRAM (A100, H100). The Lite version of HunyuanWorld 1.0 and the WAN-5B lightweight pipeline for v1.5 run on consumer GPUs with as little as 24 GB VRAM (RTX 4090). Community-quantized variants are also available.
How does HY World Model compare to Sora and Veo?
HY World Model generates real 3D scenes (not just video), is fully open-source, and supports real-time interactive control at 24 FPS. Sora and Veo 3 produce high-fidelity passive video but are closed-source, cloud-only, and do not output 3D geometry. HY World Model is #1 on Stanford's WorldScore benchmark for both video and 3D reconstruction.
Who builds HY World Model at Tencent?
The HY World Model is developed by Tencent's Hunyuan team. The broader Hunyuan AI initiative is led by Chief AI Scientist Shunyu Yao (former OpenAI researcher, appointed December 2025). In March 2026, Tencent folded its AI Lab into the Hunyuan team, consolidating AI efforts and bringing in video/3D expertise from the WizardLM team.
What components make up HY-World 2.0?
HY-World 2.0 is a four-stage pipeline: HY-Pano 2.0 generates a 360° panorama from text or image, WorldNav plans camera trajectories, WorldStereo 2.0 expands the panorama into a navigable 3DGS world, and WorldMirror 2.0 — a ~1.2B-parameter feed-forward model — predicts depth, surface normals, camera parameters, point clouds, and 3DGS in a single pass. At launch, WorldMirror 2.0 is open-sourced; HY-Pano 2.0 and WorldStereo 2.0 weights follow shortly.
Which game engines does HY-World 2.0 export to?
HY-World 2.0 outputs Mesh, 3D Gaussian Splatting (3DGS), point clouds, depth maps, and camera parameters that drop directly into Blender, Unity, Unreal Engine, and Isaac Sim. Maxon has also integrated Hunyuan 3D into Cinema 4D for production pipelines.
What is the HY-World 2.0 license?
The model and weights are released under the tencent-hy-world-2.0-community license. The full repository is at github.com/Tencent-Hunyuan/HY-World-2.0 and the model card at huggingface.co/tencent/HY-World-2.0.
What is Character Mode?
Character Mode lets you explore a generated 3D world in first or third person with physics-based collision. Unlike video-only world models, there is no fixed duration — the world is persistent 3D geometry, so playable length is unlimited.