Tencent releases Hy-Nomas interactive world-model weights
Tencent's Hy-Nomas lets text and camera controls guide video worlds over multiple turns. Apache-2.0 code and weights are public, with clear limits.
Hy-Nomas is Tencent's released interactive video world-generation model. It extends a generated world segment by segment using multi-turn text prompts or camera controls. Code and weights are publicly available under Apache 2.0; the release is not independent proof of physical simulation, continuous real-time play or robotics readiness.
CLAIM-LEVEL EVIDENCE
What is documented and what remains open
- Primary source
Code and weights shipped
Tencent's GitHub repository and Hugging Face page point to downloadable assets.
- Vendor claim
Multi-turn scene control
Text instructions, camera control and persistent history are developer-described features.
- Vendor claim
Runtime settings
24 FPS is playback output; one-RTX-4090 examples use quantization and offloading, not proven real-time speed.
- Open question
World-model fidelity
No independent evidence establishes faithful physics, long-session reliability or robotics certification.
Vendor facts do not establish independent product testing.
Background and practical implications
Most text-to-video products generate a self-contained clip. Tencent's Hy-Nomas instead aims at a world that can be steered as it unfolds. A user can provide additional instructions after one segment or change the camera trajectory while subsequent frames are generated. Its official repositories list code, model assets and an interactive demo, with the publicly released package dated October 23, 2026. This is a genuine open artifact rather than simply an announced research direction, though the scope of its practical usefulness must be evaluated separately.
The architecture is described as a segment-autoregressive diffusion transformer combined with a text encoder, camera conditioning and a video decoder. A persistent key/value cache carries history between successive segments to improve continuity. That memory mechanism can help an object or scene remain recognizable, but it is not a demonstration of perfect long-horizon consistency or a faithful physical world model. Tencent's card lists an output configuration of 720 by 1280 pixels at 24 frames per second. The frame rate describes video playback, not measured generation throughput.
For more efficient generation, the release documents an eight-step distilled sampling configuration with a compatible decoder. A separate low-memory setup uses quantized INT8 model weights, an INT8 cache, an FP8 decoder and CPU offloading to support one RTX 4090 GPU. That configuration is specific and may trade speed for memory. Developers should not extrapolate that all workflows run smoothly on consumer hardware or that multi-turn sessions can grow indefinitely without additional host memory. The project notes that archived history can expand over time.
Licensing also deserves a precise reading. Tencent places the publicly released Hy-Nomas code and model weights under Apache License 2.0, while third-party components retain their respective obligations. Commercial teams need to review all included encoders, decoders, model dependencies, safety documentation and applicable content policies rather than treating a repository badge as universal clearance. The GitHub instructions also require a compatible CUDA stack and several attention-related dependencies. Reproducibility therefore depends on more than downloading a checkpoint.
The immediate opportunity is experimental interactive video production: evolving storyboards, controllable camera studies and prototypes of persistent virtual scenes. Training embodied AI may eventually benefit from more capable world generators, but Hy-Nomas is not independently established as a trustworthy robotics simulator. Our assessment is that open weights, multi-turn control and developer access make this release important. Robust physics, safety, practical latency and quality across extended interaction remain questions for repeatable external evaluation.
Primary sources
- Tencent Hunyuan — Hy-Nomas GitHub repository
- Tencent — Hy-Nomas model card and weights
- Tencent — Hy-Nomas license
Related articles
Frequently asked questions
Are the full Hy-Nomas assets publicly available?
Tencent provides its code and model files through GitHub and Hugging Face.
Does 24 FPS mean 24 generated frames every second?
No. The number describes output video playback. Rendering time is dependent on hardware and configuration.
Is Hy-Nomas already a validated robotics simulator?
That is not established. It is an interactive generative video-world research release.