Behold: Genmo dropped Mochi 1 under Apache 2.0 — full commercial use, no asterisks, no "research only". The clips are smooth, the prompt adherence is sharp, and the timeline is finally calm about open video gen.
The Setup
Mochi 1 is a 10B AsymmDiT with a custom VAE that compresses video aggressively. The motion is what sets it apart — no morphing faces, no melting hands, just believable physics. It lives on HuggingFace and the diffusers PR is already merged.
pip install -U diffusers transformers accelerate imageio[ffmpeg]
huggingface-cli download genmo/mochi-1-previewThe Money Pattern
The pipeline is identical to any other diffusers call. Plug a prompt in, get an mp4 out, pipe it into FFmpeg for post. For motion plates and Aidxn Design portfolio loops this is my new default.
import torch
from diffusers import MochiPipeline
from diffusers.utils import export_to_video
pipe = MochiPipeline.from_pretrained(
"genmo/mochi-1-preview",
torch_dtype=torch.bfloat16,
)
pipe.enable_model_cpu_offload()
pipe.enable_vae_tiling()
frames = pipe(
prompt="aerial shot over Gold Coast skyline at sunset, slow drift, cinematic",
num_frames=85,
num_inference_steps=64,
guidance_scale=4.5,
).frames[0]
export_to_video(frames, "mochi.mp4", fps=30)The Catch
480p is the ceiling on the open weights — Genmo kept the HD variant API-only. Generation is also slow: 64 steps on a 4090 takes about 3 minutes per clip. And the prompt window is short, so wordy descriptions get truncated and ignored.
The Verdict
Apache 2.0 is the headline. You can ship Mochi-generated clips commercially, full stop. The 480p limit hurts but upscaling through Topaz or even an SD4 img2img pass gets you there. This is the foundation model the open-source video space needed. Pull it.