Mozilla's Human-Scale AI Incubator Pays Up to $500,000 for Small Models. The Application Asks Your Parameter Count
November 5, 2026 · 6 min read
Granted Research Team · Editorial policy
Most AI grant applications ask what your model can do. Mozilla's asks how big it is.
Among the screening questions on the application form for the Mozilla Foundation's Human-Scale AI Incubator, there is a dropdown: "What is the largest parameter count of any AI model core to your project's functionality?" The answer choices start at "10 billion or fewer" and end at "More than 40 billion." Just before it comes a yes-or-no question: "Does your project's core functionality depend on calling the API of a frontier AI model?" The help text names the GPT, Claude, Gemini and Llama families.
Applicants whose answers fall outside the call's focus see a message: "Based on your answers, your project falls outside the focus of this particular call for proposals." The form does not say which answers trigger it. The design makes the intent plain. Mozilla is paying for AI that does not depend on the largest labs.
The money is substantial for this kind of work. Mozilla will select seven projects. Each receives $300,000 and 12 months of support in a cohort that begins in April 2027. Projects that show strong progress can receive up to $200,000 more and two additional years of support, for a potential total of $500,000 over three years. The funding is non-dilutive: "We fund, you build, and you keep ownership." Applications close Friday, November 20, 2026, at 11:59 p.m. Pacific Standard Time (UTC−8). That is 15 days from today.
Seven Projects, Three Principles
The call is subtitled "Small Models. Intentional Systems. Tailored Solutions." Mozilla says projects must embody all three of its emerging Human-Scale AI design principles. The form turns each one into a required essay question:
| Principle | What the application asks |
|---|---|
| Small | "What models, hardware, and infrastructure does your project use, and why are they the right fit for the problem and context you are addressing?" |
| Intentional | "How do users understand and control how their data is collected, stored, and used? What agency do they or their community have over the system and its governance?" |
| Tailored | "How have the people, problem, and context you're designing for shaped the technology itself, including what it does, how it works, or how it's deployed?" |
Coverage of the call (Opportunities Feed, DEV Community) lists the themes Mozilla highlighted:
- AI that runs offline or on basic mobile phones
- Smaller models specialized for specific tasks
- Community-owned technology, including speech recognition
- Data privacy and user control
- Locally relevant solutions
- Transparent, adaptable systems
Each principle works as a filter. "Small" excludes a thin wrapper around a hosted frontier model. "Intentional" excludes systems in which users cannot see or govern their own data. "Tailored" excludes general-purpose tools that someone later tried to fit to a community.
This Is Not a Seed Grant
The second thing most applicants will miss is the stage requirement. The form says: "We're looking for working technologies already deployed with the people they're designed to serve, even at a small scale or through an early pilot." Projects must show "early evidence of product-community fit, while still having significant technical, product, community, deployment, or sustainability work ahead."
The form asks applicants to place themselves on a four-step ladder: "Concept or wireframe only," "Working prototype," "Deployed with early users (less than 100 users)," and "Deployed at scale (100+ users)." By Mozilla's own description, concept-stage projects are out of scope. Coverage of the call also says generic chatbots, standalone research or policy projects, and ideas without working technology do not fit.
The target is a narrow middle stage: past the pilot, with real users, but before the project is sustainable. Many projects stall there, because seed funders have moved on and later-stage funders want revenue. A $300,000 grant with no equity attached suits a project at that point.
Who Can Apply
The eligibility rules are broader than most foundation AI programs. The form lists six applicant types: individual, nonprofit or NGO, fiscal sponsorship, for-profit or other organization, governmental unit, and other. Mozilla says you do not need to be a nonprofit. Two conditions apply to everyone:
- You must be legally able to receive grant funds from Mozilla Foundation. The form includes a country dropdown. Applicants from ineligible locations are told Mozilla "is unable to provide funding connected to your location." The form does not list those countries, so check the foundation's grant eligibility guidelines before investing time.
- The project must serve a charitable purpose. The form asks: "Does this project serve a charitable purpose and create meaningful public benefit?" It then asks who benefits "beyond your organization." A for-profit startup can apply, but it needs to describe a public benefit that does not depend on its own revenue model.
For companies, this is a rare source of non-dilutive AI funding that does not require a federal SBIR application. Before applying, founders should confirm with counsel how a charitable-purpose grant can be used, since foundation grants to for-profits generally carry use restrictions and reporting.
How It Compares to Mozilla's Earlier Cohort
Mozilla's previous incubator round, the Democracy x AI Cohort, opened in February 2026. It funded 10 projects at $50,000 each, and two finalists could receive an additional $250,000 through a Sustain Track. That design spread small grants widely and concentrated follow-on money on two winners.
The Human-Scale AI cohort reverses that design. Fewer projects get much more money at the start: seven grants of $300,000 total $2.1 million in year-one awards. If every project earned the full follow-on, the cohort would total $3.5 million. That is our arithmetic, not a Mozilla figure. The change in structure signals that Mozilla is backing teams it expects to last, not running a pipeline test.
How to Write a Competitive Application
The form tells you what reviewers will look for. Use it.
Answer the parameter question as a design choice. If your model is a 3-billion-parameter speech model running on a $60 phone, that is the core of your case. Explain in the "Small" essay why a larger model would be worse for your users, not just more expensive: latency, connectivity, cost to the community, or data that cannot leave the device. If you do call a frontier API, say exactly where, and why the system's core does not depend on it.
Count your users precisely. The form asks for traction in user bands ("Less than 100" up to "100,000+") and for evidence of fit: users, contributors, community, partners and downstream adoption. "You don't need all of these," it says. Concrete numbers from a real deployment will beat a large projected market.
Use the governance essay well. The "Intentional" question asks what agency users or their community have "over the system and its governance." The strongest answers describe mechanisms, such as data cooperatives, community review boards, local hosting and opt-in training. Privacy policy language will not carry this section.
Explain what the $300,000 makes possible. The form asks, "What would this $300,000 grant unlock for your project?" and also asks about 12- and 36-month success measures and long-term sustainability. Reviewers will check whether the year-one plan builds toward the $200,000 follow-on.
Work in the open. The form asks how you will share code, learnings and data. Open-sourcing is not stated as a requirement, but it is clearly favored.
Record the video. A two-minute video showing the project and team is "optional but strongly encouraged." For a call that requires deployed technology, a video of the system working in the field is the most persuasive evidence you can submit.
Write it yourself. The form asks whether AI or translation tools were used. It warns that applications "that appear to be heavily or entirely AI-generated, or that do not demonstrate authentic engagement with the application questions, may be subject to further review." Use tools to organize and edit, but make sure the substance, numbers and voice are clearly your team's.
Applications are accepted in English, German, Spanish, French, Polish and Portuguese. The cohort is meant to connect builders "across communities and countries," so projects outside the United States are clearly in scope, subject to the location check.
With 15 days left and seven slots worldwide, applicants whose deployment data is ready and whose small-model choice is deliberate will have the advantage.
If you are turning a working deployment into a competitive application, Granted can help structure your traction evidence and draft the principle essays before the November 20 deadline.