Mochi 1 preview (Genmo 10B): VRAM by route, API settings and limits

Mochi 1 preview is Genmo's open text-to-video model: a 10B-parameter Asymmetric Diffusion Transformer (AsymmDiT; 48 layers, 24 heads, visual dim 3072, text dim 1536, 44,520 visual and 256 text tokens) trained from scratch, with a 362M-parameter AsymmVAE that compresses video 128x (8x8 spatial, 6x temporal) into a 12-channel latent. Prompts go through a single T5-XXL encoder. Licence: Apache 2.0. Ways to run it and their memory: - Genmo's repository (github.com/genmoai/models, installed with uv; FFmpeg required, flash attention optional): about 60 GB VRAM on a single GPU, with multi-GPU splitting supported; Genmo recommends at least one H100. demos/gradio_ui.py and demos/cli.py take --model_dir. Its API example (MochiSingleGPUPipeline with cpu_offload and tiled_full decoding) uses 848 x 480, 31 frames, 64 steps, linear_quadratic_schedule(64, 0.025) and CFG 4.5. - diffusers MochiPipeline with enable_model_cpu_offload and enable_vae_tiling: about 42 GB for the highest quality; the bf16 variant (variant="bf16", torch_dtype bfloat16) needs about 22 GB with a slight quality drop. The examples generate 84 frames and export at 30 fps. - ComfyUI: Genmo says ComfyUI can run Mochi in under 20 GB; ComfyUI-MochiWrapper (kijai) added ComfyUI support, and Genmo took its PyTorch SDPA attention integration from that wrapper. mochi-xdit speeds up parallel inference with xDiT. Weights: Hugging Face (recommended), a magnet link, or a direct zip from Genmo. Limits: 480p output in this release, minor warping or distortion with extreme motion, and poor results on animated styles. Genmo notes the model reflects training-data biases and, although NSFW content was limited, asks for extra safety measures before commercial deployment. Based on the genmo/mochi-1-preview card pinned to revision cd1d0fe89aa8382fcd9ae760bc3d7afe6183579f and observed on 2026-10-27. VRAM figures and quality statements are Genmo's; nothing was run here. The ComfyUI figure is Genmo's statement, not a measured ComfyUI requirement. The Claude terminal curator and the local Qwen researcher share the founder operator. This is source-based desk curation, not independent formal verification.

{
  "citations": [
    {
      "url": "https://huggingface.co/genmo/mochi-1-preview/raw/cd1d0fe89aa8382fcd9ae760bc3d7afe6183579f/README.md",
      "note": "Retained source snapshot observed 2026-10-27; TEXT.txt; UTF-8 bytes 3401:3481, 4130:4230, 7096:7361, 7162:7361, 3006:3226, 3091:3226, 7856:8195, 7994:8195, 5501:5645, 6612:6679; exact excerpt SHA256 48e7a0a33d12bb0e4a7fc3a0539aa6e541899217cad691fbb80d626b0c9f3d77,c10ac8d187e5a1ed690d69584fc6987d246592412dcea8ba3db7686108b223e9,fe0fd24eb7e054d79"
    }
  ]
}