On December 30, 2025, the Tencent Hunyuan team officially open-sourced HY-Motion 1.0 (Hunyuan-Motion-1.0), a large-scale text-to-3D motion generation model with ten billion parameters. Based on the Diffusion Transformer (DiT) architecture and flow matching mechanism, this model can generate high-fidelity, smooth, and diverse 3D character skeleton animations with just one natural language description. It directly supports mainstream 3D tools such as Blender, Unity, and UE, significantly lowering the barrier to animation production.
Key Technical Highlights
HY-Motion 1.0 adopts a full-stage training strategy: first pre-training on over 3,000 hours of diverse motion data to build general motion priors; then fine-tuning on 400 hours of high-quality selected data to improve detail smoothness; finally, optimizing physical plausibility and semantic alignment through reinforcement learning (RLHF) combined with human feedback and reward models.
The model covers six major categories with over 200 types of motions, including basic movements, sports competitions, fitness and outdoor activities, social leisure, daily activities, and game character actions (such as sword blocking and zombie walking). The output is in SMPL-H skeleton format, supporting atomic actions, composite sequences, and concurrent actions.
Outstanding Test Performance
Community tests show that the model has a high level of accuracy in daily scenarios: for prompts like "running," "sitting on a chair," and "jumping twice with both legs," the generated motions are natural and continuous. Complex motions, such as crouching during the bullet time scene from "The Matrix," can also be accurately reproduced with smooth postures.
In performance evaluation, the model achieved an instruction following capability of 78.6% (SSAE metric), with an average action quality of 3.43 out of 5, surpassing open-source baselines such as MoMask and DART, especially in complex instructions and multi-category coverage.
Extreme challenges: the recreation of professional athlete movements (such as ski jumping, diving, and BMX cycling) is not ideal, and joint transitions occasionally appear unnatural. However, overall physical plausibility far exceeds previous versions.
Great Potential in Game Animation Applications
This end-to-end model is particularly suitable for game development: it can quickly generate NPC daily life actions (such as walking and interaction), significantly accelerating research and iteration. Although main character design requires later refinement, it can already be seamlessly imported into engines, helping to create MMORPGs and action games. Film storyboarding, advertisement positioning, and VR content creation will also benefit.
The lightweight version HY-Motion-1.0-Lite (0.46B parameters) is also open-sourced, offering more friendly deployment.
Project link: https://hunyuan.tencent.com/motion?tabIndex=0
