HYMotion-G1
This is the self-contained Motius release of the HYMotion-G1 339k checkpoint. It includes the 0.46B motion transformer, G1-38 normalization statistics, classifier-free guidance embeddings, Qwen3-8B, and CLIP-L.
from motius import Pipeline
pipe = Pipeline.from_pretrained(
"ZeyuLing/Motius-HYMotion-G1",
device="cuda",
)
result = pipe.infer_text_to_motion(
"a robot walks forward and waves",
num_frames=180,
)
g1_38 = result["g1_38"]
qpos_36 = result["g1_qpos"]
The native output is canonical g1_38 at 30 fps. The g1_qpos decode is
exact and does not pass through SMPL or inverse kinematics. See the
Motius Model Card
for training, representation, and benchmark details.