HY-Motion 1.0 is Tencent Hunyuan's open release of a text-to-3D human-motion generation system: given a plain-language prompt, it produces skeleton-based 3D character animations that can be dropped into standard animation pipelines. It is built on a Diffusion Transformer (DiT) with Flow Matching, and its headline claim is that it is the first project to scale DiT-based text-to-motion models to the billion-parameter level. Released with pretrained weights, a Hugging Face Space demo, and an arXiv technical report, it targets a persistent bottleneck in games, film, and virtual production: animating characters is slow and specialized work, and generating natural motion from text could remove much of that friction. ## What It Is HY-Motion 1.0 turns a text description into a sequence of skeletal poses — the motion of a character's joints over time — rather than a rendered video or a static mesh. That choice is deliberate and important: because the output is skeleton-based, it is directly compatible with the rigs and retargeting tools that 3D artists already use, so a generated walk, gesture, or fight sequence can be applied to an existing character and refined like any other animation. The model is offered in a standard 1.0-billion-parameter version and a lightweight 0.46-billion variant, HY-Motion-1.0-Lite, for teams with tighter compute budgets. ## Scaling DiT to a Billion Parameters The technical thesis is that text-to-motion generation has been held back by small models, and that scaling a Diffusion Transformer with Flow Matching to the billion-parameter range yields markedly better instruction following and motion quality. In the project's evaluations it reports state-of-the-art results against comparable open-source models on both how faithfully generated motion matches the prompt and how natural the motion looks. Flow Matching provides an efficient generative formulation for the continuous, temporally coherent trajectories that motion requires, and the DiT backbone supplies the capacity to learn a broad range of actions and styles. ## Three-Stage Training HY-Motion's quality comes from a deliberate three-stage training recipe. It begins with large-scale pretraining on over 3,000 hours of diverse motion data to learn a broad motion prior, then fine-tunes on roughly 400 hours of curated, high-quality 3D motion to sharpen detail and smoothness, and finally applies reinforcement learning from human feedback and reward models to further improve instruction following and naturalness. This progression — broad prior, quality refinement, preference alignment — mirrors the training pipelines that have proven effective for large language and image models, adapted here to the motion domain. ## Model Zoo and Access Tencent ships the model openly: pretrained weights for both the standard and Lite variants are on Hugging Face, there is an interactive Hugging Face Space so the model can be tried in a browser before any local setup, and a companion arXiv report documents the architecture and training. The standard 1.0B model lists a minimum VRAM requirement of about 26GB, which situates it on data-center or high-end workstation GPUs, while the Lite variant lowers the barrier for smaller setups. An official project site and demo round out the release. ## Usability in Practice For technical artists and researchers, HY-Motion is unusually easy to evaluate because of the hosted Space: a prompt in, a motion clip out, with no installation. Integrating it into a real pipeline is more involved but conceptually familiar, since the skeleton-based output aligns with existing rigging and retargeting workflows. The honest caveats are the hardware bar — 26GB of VRAM for the standard model puts local use out of reach for most consumer GPUs, though the Lite model helps — and the license: HY-Motion is released under the custom Tencent Hunyuan Community License rather than a permissive MIT or Apache grant, so teams must review the terms before commercial deployment. As a 1.0 release focused on human motion, it also does not attempt to cover non-humanoid or highly stylized motion beyond its training distribution. ## Pros and Cons The strengths are a genuinely useful task framing — skeleton-based motion that fits existing pipelines — strong reported instruction following from billion-scale DiT modeling, a rigorous three-stage training recipe including RLHF, two size options for different compute budgets, and an open release with weights, a live demo, and a technical report. The trade-offs are the high VRAM requirement for the standard model, a custom community license rather than a standard open-source one, a scope limited to human motion, and the usual gap between benchmark results and the messiness of production animation needs. ## Outlook HY-Motion lands as generative 3D moves from static objects toward motion and behavior, the parts of a scene that are most labor-intensive to author by hand. A capable, openly released text-to-motion model from a major lab is a meaningful step, and the billion-parameter scaling result suggests headroom for further gains as data and model size grow. If subsequent releases broaden beyond human motion and ease the hardware and licensing constraints, the approach could become a standard part of animation toolchains. ## Conclusion HY-Motion 1.0 is a serious, openly released text-to-3D-motion model that generates pipeline-ready character animation from text. It is most valuable to game and film studios, technical artists, and motion-generation researchers who can work within its hardware and licensing terms, and less suited to hobbyists on consumer GPUs, for whom the Lite variant and hosted demo are the natural entry points. For text-driven 3D character animation in 2026, it is one of the most capable open options available.