Genmo
Genmo is a San Francisco research lab whose Mochi 1 model turns written prompts into short videos. Generate 480p MP4 clips of up to 5.4 seconds in its web playground, or run the Apache 2.0 weights locally.
What is Genmo?
Genmo presents itself less as a video app than as a research lab, "dedicated to building open, state-of-the-art models for video generation towards unlocking the right brain of AGI". Its stated ambition is to develop "the world's most sophisticated video world models to understand the physical world like never before".
The product carrying that ambition is Mochi 1, a text-to-video diffusion model released as a research preview under the Apache 2.0 licence and described by the company as the largest video generative model ever openly released. It runs on a novel Asymmetric Diffusion Transformer (AsymmDiT) architecture, trained entirely from scratch, with 10 billion parameters. The visual stream carries roughly four times more parameters than the text stream, prompts are encoded by a single T5-XXL model where other diffusion systems stack several encoders, and non-square QKV projections keep inference memory down. The model reads a 44,520-token video context window with full 3D attention and RoPE extended to three dimensions, and borrows SwiGLU, query-key normalisation and sandwich normalisation from LLM scaling work. An open video VAE compresses input 96x, eight by eight spatially and six temporally, into a twelve-channel latent space. Genmo evaluates two axes: prompt adherence, judged by a vision-language model following the DALL-E 3 protocol, and motion quality, scored with Elo ratings on the LMSYS Chatbot Arena protocol. It claims fluid simulation, believable fur and hair, and human motion beginning to cross the uncanny valley.
Two routes lead to the model. The hosted playground requires an account, applies safety moderation to generations, and runs on a credit system called fuel, at 100 credits per Mochi video and 50 for a Replay. A generation takes two to five minutes and returns an MP4 (H.264) clip of up to 5.4 seconds at 30 fps, 480p on the base model. The open route is a download: weights on Hugging Face, code on GitHub, execution through ComfyUI or a Python command line.
The limits are acknowledged by the publisher: 480p output, occasional warping and distortion under extreme motion, and weak results on animated styles, the model being optimised for photorealism. The team behind it comes from the DDPM, DreamFusion and Emu Video work, with more than 50,000 cumulative citations claimed.
What it does
- Generate a video from a written description, in plain language
- Produce MP4 H.264 clips of up to 5.4 seconds at 30 fps, 480p on the base model
- Start from a random prompt suggestion when inspiration runs short
- Download the finished videos, which stay available for 30 days
- Browse and share community work through the Discover and Creations tabs
- Download the Mochi 1 weights from Hugging Face and run the model on your own machine
- Customise or fine-tune the model through the GitHub repository or ComfyUI
When to use Genmo / When not to
A quick filter to help you decide if Genmo is the right fit.
When to use Genmo
- Artists and video creators who want to turn a written idea into a short clip without any technical skill
- Marketing and social media teams producing short photorealistic assets, on a paid tier that authorises commercial use
- Computer vision and video generation researchers looking for open Apache 2.0 weights and a deliberately hackable architecture to study or fine-tune
- Robotics and autonomous systems teams generating synthetic video data to train perception models
- Product teams prototyping generative video features for entertainment, advertising or education applications
When not to use Genmo
- Anyone who needs long-form video: clips are capped at 5.4 seconds
- Broadcast or HD work: the base model outputs 480p, and the HD version announced on the blog was not available at the time of review
- Animation and cartoon production: the publisher states the model is optimised for photorealistic styles and does not perform well with animated content
- Developers who need a documented public API or image-to-video, neither of which the site offers today
- Organisations with strict confidentiality requirements: user data feeds model training, with no documented opt-out and no published data processing agreement
How to use Genmo
A typical end-to-end flow, from setup to results.
- Create an account, or sign in with Google or Discord
- On the free tier, add a payment method to unlock the 250 lifetime credits; it is never charged
- Open the Mochi 1 playground from the site's Play page
- Write a descriptive prompt in plain language, using the prompt ideas helper or the random prompt generator if needed
- Click Generate
- Wait two to five minutes, depending on prompt complexity and server load, with progress shown in real time
- Download the finished video; it stays available for 30 days
- Browse the Discover and Creations tabs to see and share community work
- Move to a paid tier for watermark-free output, commercial use and higher queue priority
- Developer route: clone the Mochi repository from GitHub, install the package and run the Python demo command line, or load the Hugging Face weights into ComfyUI
Pros & Cons
Pros
- Weights and code are open under the Apache 2.0 licence, for personal and commercial use
- The model can be run on your own hardware, so there is no compulsory dependency on the hosted service
- A free tier gives 250 lifetime credits; a payment method is required to unlock them but is never charged
- No technical skill is needed on the hosted route, a plain-language prompt being enough
- The architecture is publicly documented and presented as simple and hackable
- The research team is identifiable and traceable, coming from DDPM, DreamFusion and Emu Video with a claimed 50,000-plus citations
- Financial backing is stated: a 28.4 million USD Series A led by NEA, with the service reported as fully operational
Cons
- Paid tier prices are not readable in the page source: the pricing page loads them in JavaScript, so no amount can be checked without a browser
- Clips are capped at 5.4 seconds and the base model outputs 480p
- Image-to-video is announced as coming soon and is not available
- The model is optimised for photorealism and is stated to perform poorly on animated content, with possible warping under extreme motion
- The licence granted to Genmo over Inputs and Outputs is very broad, perpetual and sublicensable, and user data feeds model training with no documented way to opt out
- No public API is documented on the site, and the free tier carries a Genmo watermark with commercial use excluded
- Signals of stale upkeep: a privacy policy dated January 2023, a single blog post, an HD version promised for 2025 and not observed, and no robots.txt or sitemap served
Pricing & Plans
A permanent free plan is offered at 0 USD per month, granting 250 lifetime credits unlocked by adding a payment method that is never charged, with a Genmo watermark on the generated videos. Two paid tiers, Lite and Standard, are advertised, but their amounts are not published in the page source: they are rendered in JavaScript and still displayed as "Loading..." at the time of review, so no entry-level paid price can be stated here. Billing is in US dollars, with a monthly or annual choice and 20% off the annual commitment. There is no time-limited free trial: what exists is a permanent free tier.
- 0 USD per month
- 250 lifetime credits
- Genmo watermark on the generated videos
- 1
- 200 credits per month
- no watermark
- commercial usage and high queue priority
- the amount is not published in the page source
- 5
- 000 credits per month
- no watermark
- commercial usage
- top queue priority and early access to new models
- the amount is not published in the page source
- A Pro plan is mentioned in the site's FAQ but was not displayed on the pricing page at the time of review
- Billing is monthly or annual
- with 20% off the annual commitment
- credit balances are topped up each month
- a Mochi video costing 100 credits and a Replay video 50
- Subscriptions renew automatically
- cancellation must be made at least three days before the renewal date
- and no pro-rata refunds are given
Data, GDPR & hosting
A consolidated view of how Genmo handles your data.
GDPR overview
Genmo makes no claim of GDPR compliance anywhere on the site, and the word GDPR does not appear once. What exists is thin. The privacy policy tells people in the EEA or the UK that they may complain to their local supervisory authority, links to the European Commission's directory, and gives the Swiss regulator's details. It states a right to withdraw consent at any time without retroactive effect, plus rights of access, correction and deletion "based on the applicable laws of your country", exercised through a Termly-hosted form or at hi@genmo.ai. No Article 27 EU representative is designated, no data protection officer is named, no DPA is published, and no legal bases, standard contractual clauses or transfer mechanisms are described. The terms state the service is controlled and operated from the United States; the privacy policy is dated 19 January 2023.
Who owns the data?
Under the terms of service you keep any copyright and proprietary rights you hold in your Inputs and Outputs, and Genmo claims no ownership of Outputs. In exchange you grant Genmo a worldwide, royalty-free, non-exclusive, perpetual, irrevocable, transferable, assignable and fully sublicensable licence, through multiple tiers and without attribution or payment, to host, store, use, exploit, reproduce, distribute, publicly display, adapt, modify and create derivative works from that material, in any format, including to build, train, operate and improve its services, products and technologies. Moral rights are waived, the licence survives termination, and Genmo may remove or edit Inputs and Outputs without notice. Feedback is treated as non-confidential and freely usable.
Reuse rights
Outputs produced in the playground may not be used for commercial purposes unless they were generated during an active paid subscription that expressly authorises such use; the terms otherwise reserve the service to personal, non-commercial use. Consistently with that, the pricing page carries a Commercial usage line on the Lite and Standard tiers only, while the Free tier adds a Genmo watermark. Videos can be downloaded and stay available for 30 days after generation, and export or sharing tools point to third-party services, with Genmo disclaiming any responsibility for downstream use. The terms give no warranty that you legally own an Output or that it does not infringe third-party rights. The Mochi 1 model itself is a separate question: released under the Apache 2.0 licence, it is described as freely available for personal and commercial use, which covers the weights and the code rather than the videos made on the hosted service.
Data retention & training
Hosting summary
The site declares no hosting jurisdiction. Nothing in the terms, the privacy policy or the FAQ names a data centre, a region or a cloud provider for the storage of user data. What is stated is legal rather than technical: the terms say the services are controlled and operated from the United States, apply the law of the State of California and of the United States, and give exclusive jurisdiction to the courts of San Francisco County. The privacy policy lists categories of third parties that may receive data, namely cloud computing services, data analytics services, data storage providers, performance monitoring tools and content moderation services, and names a single entity, Stability.ai. It announces technical and organisational security measures without citing any standard or certification: no SOC 2, ISO 27001 or HIPAA claim appears anywhere, and no international transfer mechanism or standard contractual clauses are described. For information only, and not as a publisher commitment: the domain resolves to 76.76.21.21, an address announced by AS16509 (Amazon.com Inc.) and geolocated in the United States, behind a Vercel front end.
Things to keep in mind
Risks and trade-offs to weigh before adopting Genmo.
- Every prompt you type and every video you generate falls under a perpetual, irrevocable and sublicensable licence granted to Genmo, which survives the end of your account
- The information you supply and the images you generate are used to improve the models, with no documented way to opt out; typed prompts are explicitly listed among the personal data collected
- Other users may obtain and exploit outputs similar or identical to yours from comparable prompts, and the terms give no warranty that you legally own an output or that it infringes nobody's rights
- The publisher's maximum liability is capped at 50 USD, disputes go to mandatory arbitration with a class action waiver, and opting out of arbitration requires a postal letter sent within 30 days
- Generative video carries generic misuse risks, from deepfakes to misinformation and impersonation of a person's appearance, and the site mentions no provenance marking on the videos it produces
- Generated videos stay downloadable for only 30 days, and an account may be suspended or terminated at any time without notice, with immediate loss of access and no refund
- Easy generation invites over-reliance: a research preview with acknowledged limits can quietly replace storyboarding, shooting and editing work, and a generation can also simply fail on the content policy
Setup & Integrations
Technical difficulty
Two very different routes. The hosted playground is easy: sign up with Google or Discord, add a payment method to unlock the free credits, which is never charged, then type a prompt. Nothing to install or configure, and the FAQ states that no technical skills are required. Running Mochi 1 yourself is another matter: cloning the GitHub repository, installing the package and driving the Python command line, or loading the Hugging Face weights into ComfyUI, means handling a 10-billion-parameter model on your own GPU, and the site publishes no minimum hardware specification.
Deployment
Integrations
Behind Genmo
Fundraising
Social
Resources
All the official URLs gathered for verification and reference.
Frequently asked questions
Which model does Genmo run on?
How long and how large are the generated videos?
How long does a generation take?
Do I need technical skills to use it?
Is there a free plan?
What is fuel?
Can I use the videos commercially?
Is my data used to train the models?
Can I run the model myself?
Is there an API, and is image-to-video available?
Should you pick Genmo?
Genmo is best read as what it says it is: a research lab that publishes its work openly, rather than a polished video SaaS. Its single product, Mochi 1, is a 10-billion-parameter text-to-video model released under the Apache 2.0 licence, with weights on Hugging Face and code on GitHub. That gives the tool two lives: a hosted playground with a free credit tier, and a local install with no dependency on the company at all. This openness is the strongest reason to look at it, and it is uncommon at this parameter count.
The natural audience is creators and marketing teams making short photorealistic clips, computer vision researchers who want a documented and deliberately hackable architecture, and robotics teams producing synthetic training data, a use case Genmo names itself.
The reservations are real. At the time of review, output was capped at 5.4 seconds and 480p on the base model, image-to-video was still announced as coming soon, and no public API was documented. Contractually, the terms grant Genmo a perpetual, irrevocable and sublicensable licence over every prompt entered and every video produced, user data feeds model training with no documented opt-out, and commercial use of playground output requires a paid subscription that authorises it. The prices of those paid tiers are not readable in the page source, which makes budgeting impossible without opening the site in a browser.
Signals of solidity exist: a 28.4 million USD Series A led by NEA, a research team with a traceable publication record, a service reported as operational. So do signals of drift: one blog post, an HD version promised for 2025 and not observed, and a privacy policy last dated January 2023. Worth trying, with the licence terms read first.
- Choosing a selection results in a full page refresh.
- Opens in a new window.