What is HY-World 1.5 (WorldPlay)?
HY-World 1.5, also called WorldPlay, is Tencent's open real-time interactive world model, released on 17 December 2025. Given one image and a text description, it generates video of a world that you move through with keyboard and mouse, predicting 16 frames at a time. Unlike Tencent's later HY-World 2.0, which builds 3D scenes, WorldPlay is a streaming video model.
The main model is built on HunyuanVideo 1.5 (about 8 billion parameters); a lighter 5B version on Wan followed on 6 January 2026, with lower quality. Tencent released the full training code the same day and its reinforcement-learning post-training code, WorldCompass, on 8 March 2026. The paper reports 720p at 24 frames per second; the public checkpoints are 480p image-to-video only.
The weights use the Tencent HY-WorldPlay Community License, which does not apply in the EU, the UK or South Korea and requires a separate licence above 1 million monthly active users.
Is HY-World 1.5 (WorldPlay) worth it?
WorldPlay's distinction is openness of method rather than ease of use: Tencent published the data pipeline, pre-training, reinforcement-learning post-training and distillation code, which almost no other world-model lab has done. If you want to train your own interactive world model, this is the most complete starting point.
It also tackles the problem that ruins most video world models: forgetting. A memory system rebuilds context from earlier frames, including long-past ones, so a place you return to looks the same. In Tencent's own table, long-horizon PSNR (a measure of how closely regenerated views match) is 18.94, against 9.57 for Matrix-Game 2.0 and 10.09 for Tencent's earlier GameCraft.
The catch: hardware and licence. The README lists 28 GB of GPU memory per card when split across 8 GPUs and 72 GB on a single GPU for the distilled model, so real-time play means data-centre cards. The released checkpoints are 480p, below the paper's 720p. The licence excludes the EU, UK and South Korea outright and caps free use at 1 million monthly users. Its numbers are Tencent's own, and the distilled model with RL post-training is still listed as to be released. Adoption is modest: 1,627 GitHub stars and 493 downloads last month on Hugging Face.
Our take: for a research group outside Europe, WorldPlay is the textbook; for anyone who wants to put a player in front of a generated world on one consumer GPU, Waypoint 1.5 or Matrix-Game 4.0 gets there with fewer strings attached.
Where it earns its place
A university lab working on world models
Train a domain-specific interactive world model without writing the pipeline from scratch
Start from the released 8B training code and WorldCompass RL recipe; Tencent lists 60 GB of GPU memory per card for training at 8-way sequence parallelism.
A technical artist at a games company in Asia
Turn concept art into a walkable preview for a pitch
Feed a concept image and prompt to the distilled autoregressive model on a multi-GPU server, or use Tencent's online demo; output is 480p video, fine for a mood piece, not for production.
Before you commit
- Setup
- Clone Tencent-Hunyuan/HY-WorldPlay, download HunyuanVideo 1.5's 480p image-to-video base and the WorldPlay checkpoints, then run generate.py with torchrun; prompt rewriting needs a separate vLLM server.
- Integrations
- HunyuanVideo 1.5 base; FastVideo Diffusers ports exist; Tencent hosts an online demo.
- Data and privacy
- Self-hosted, nothing leaves your servers. Tencent HY-WorldPlay Community License: no use in the EU, UK or South Korea; above 1 million monthly active users you need Tencent's permission.
- Cost at scale
- The distilled model needs 28 GB per GPU across 8 GPUs, or 72 GB on one, per session.
- Limits
- 480p public checkpoints; image-to-video only; RL-distilled checkpoint not yet released; benchmarks self-reported.
What’s good
- Full training, RL post-training and distillation code released
- Long-horizon memory: 18.94 PSNR against 9.57 for Matrix-Game 2.0 (Tencent's test)
- Keyboard and mouse control, first or third person
- Lighter 5B variant for smaller GPUs
What’s not
- Licence excludes the EU, UK and South Korea
- 72 GB on a single GPU, or 8 cards at 28 GB each
- Public checkpoints are 480p, not the paper's 720p
- All benchmark figures are Tencent's own
Best for
Research teams outside the EU, UK and South Korea who want to train or study a real-time world model end to end.
HY-World 1.5 (WorldPlay) alternatives
HY-World 1.5 (WorldPlay): common questions
- Is HY-World 1.5 (WorldPlay) worth it?
- WorldPlay is the open world model to learn from, with every stage of training published, but you need a server rack to play it and a lawyer to ship it. We rate it 3.7 out of 5.
- How much does HY-World 1.5 (WorldPlay) cost?
- Free weights and code under Tencent's community licence (not valid in the EU, UK or South Korea). You pay for the GPUs.
- What are the downsides of HY-World 1.5 (WorldPlay)?
- Licence excludes the EU, UK and South Korea; 72 GB on a single GPU, or 8 cards at 28 GB each; public checkpoints are 480p, not the paper's 720p; all benchmark figures are Tencent's own.
- Who is HY-World 1.5 (WorldPlay) best for?
- Research teams outside the EU, UK and South Korea who want to train or study a real-time world model end to end.
- How much does HY-World 1.5 (WorldPlay) cost for a team or in production?
- The distilled model needs 28 GB per GPU across 8 GPUs, or 72 GB on one, per session.
- Is HY-World 1.5 (WorldPlay) safe for company data?
- Self-hosted, nothing leaves your servers. Tencent HY-WorldPlay Community License: no use in the EU, UK or South Korea; above 1 million monthly active users you need Tencent's permission.
- What are the best HY-World 1.5 (WorldPlay) alternatives?
- The closest alternatives we have reviewed are HY-World 2.0, Matrix-Game 4.0, Waypoint 1.5 and LingBot-World 2.0 (Infinity).
Answers checked Nov 7, 2026.