Mochi
genmoai/mochiReviewed by AI
Checked by RepoPicks Radar on
What is Mochi?
The code and the open weights of Mochi 1, a text-to-video model from the company Genmo: you type a prompt and it generates a few seconds of 480p video. It comes with a simple web page you start yourself, a command line tool, a Python API and a LoRA trainer for your own videos. It is aimed at people comfortable with a terminal: you install Python 3.11 with uv, install FFmpeg, download the model weights from Hugging Face, and the README asks for roughly 60 GB of video memory on one card.
Mochi is free and open source (Apache-2.0) and can replace Runway, Kling, and Pika.
How easy is it to install Mochi?
D · terminal, several tools or a specific graphics card · Linux
Linux is the documented path and it is a D: you clone the repository, install Python 3.11 with uv, install FFmpeg, run `python3 ./scripts/download_weights.py weights/` and then `python3 ./demos/gradio_ui.py` to get a local web page. The commands are Linux shell commands (`source .venv/bin/activate`), and the hardware note recommends "at least 1 H100 GPU", with flash-attn and A100/H100 fine-tuning, so it is tied to NVIDIA cards far bigger than a home computer's.
Coming soon: install in 1 click with the RepoPicks app.
Mochi on GitHub
- Stars
- 3,765on Nov 10, 2026
- Downloads
- —all releases
- Growth
- +4stars in 6 days
Stars over time
Facts
- Runs on
- Linux
- You need to install
- Python, NVIDIA graphics card, FFmpeg
- To build from source
- Python, Graphics card, FFmpeg
- Disk space
- 950 MB–4.3 GB · estimate
- Accounts or keys
- None
- Cost
- Free · Code and weights are Apache-2.0, with no key and no fee. Genmo also runs its own playground with the same model, and renting a suitable graphics card by the hour is a real cost.
- License
- Apache-2.0 (Permissive)
- Can replace
- Runway, Kling, Pika, LTX Studio, Veo 3
- Last change on GitHub
- Oct 22, 2026
Mochi risk signals (7)
- Attentionvery high hardware requirement
README.md "Hardware Requirements": "it requires approximately 60GB VRAM when running on a single GPU. While ComfyUI can optimize Mochi to run on less than 20GB VRAM, this implementation prioritizes flexibility over memory efficiency. When using this repository, we recommend using at least 1 H100 GPU." LoRA fine-tuning needs "one H100 or A100 80GB GPU".
AI review
- Attentiongpu path not qualified in this snapshot
README.md installation section: "This security refresh validates CPU media helpers and dependency resolution. GPU inference and fine-tuning need qualification before deployment." The News section stops at "November 26, 2024", so the recent work is on dependencies rather than new model features.
AI review
- Attentionpip install skips security override
README.md: "Use the uv project install above: it applies the Pillow 13 override for MoviePy and Gradio. A plain `pip install -r requirements.txt` does not apply that override." pyproject.toml explains why: "MoviePy and Gradio cap Pillow below the security-fixed release."
AI review
- Infoweights downloaded outside github
README.md offers the weights from Hugging Face (`genmo/mochi-1-preview`), a direct zip at `weights.genmo.dev/weights.zip` and a BitTorrent magnet link (`magnet:?xt=urn:btih:441da1af...`). The repository has no GitHub releases, so nothing is published or versioned there.
AI review
- Inforesearch preview quality limits
README.md "Limitations": "Mochi 1 is a living and evolving checkpoint... The initial release generates videos at 480p today. In some edge cases with extreme motion, minor warping and distortions can also occur. Mochi 1 is also optimized for photorealistic styles so does not perform well with animated content." The API example in the README generates 31 frames.
AI review
- Infonsfw filtering is partial
README.md "Safety": the models "inherently reflect the biases and preconceptions found in their training data. While steps have been taken to limit NSFW content, organizations should implement additional safety protocols and careful consideration before deploying these model weights in any commercial services or products."
AI review
- Infocompany backed with hosted service
README.md links to Genmo's blog, its "[Playground](https://www.genmo.ai/play)" and its "[Careers](https://jobs.ashbyhq.com/genmo)" page. The same company hosts a version of the model you can also run yourself.
AI review
Evidence is shown in its original language.
What Radar could not check about Mochi
- Whether GPU generation still works with these dependency versions, which the README itself says needs qualifying
- How good the videos look and how long each one takes on a given card
- How much disk space the weights really take, since the README lists only dit.safetensors and decoder.safetensors without sizes and the checks could only estimate
- Whether it can be made to run on a card with less than 60 GB of video memory without the ComfyUI version
- Whether it runs on Windows or Mac, which the project never mentions
- What the files behind the direct zip and the magnet link contain