What Digital Promise is funding
Digital Promise has opened a request for proposals for up to $8 million to build an open-source AI model for K-12 math tutoring in the United States, drawing on a grant managed by the Gates Foundation. The initiative, referred to as EDU AI, sits inside the organization's K-12 AI Infrastructure Program. Applications are due July 31, 2026, with work expected to begin in November 2026 and a grant period estimated at 30 to 36 months. The RFP seeks teams that combine AI engineering, classroom experience, learning science, and edtech product partnerships.
The funded deliverables are explicitly infrastructure, not a consumer app: model weights, training and fine-tuning code, datasets where permissible, evaluation tools, testing harnesses, model cards, documentation, and a reference implementation that other developers and researchers can build on. Everything must be released under open licenses. The RFP specifies at least Creative Commons Attribution 4.0 for content and Apache 2.0 or a similarly permissive license for software, code, and models, terms negotiated during the pre-award process.
The larger program behind the grant
The tutoring RFP is one piece of a $26 million, four-year program that plans to issue roughly 30 grants to develop openly shared datasets, models, benchmarks, and other digital public goods for AI in education. The program is led by Digital Promise with partners including Learning Data Insights, DrivenData, the Massive Data Institute at Georgetown University, and Catalyst at Penn GSE. Its premise is that the accuracy and relevance of classroom AI depends on shared infrastructure that no single vendor owns.
The program has already named its first grantees. On June 29, 2026 it funded four teams: Learning Equality, Princeton University, a National Tutoring Observatory effort at Cornell University, and Stanford University, working on projects from science-misconception benchmarks to a multimodal knowledge-building dataset for formative assessment. Those grants run 6 to 12 months each, a fast cadence designed to produce usable public goods rather than long research cycles. John Whitmer, a senior researcher at Learning Data Insights, said the selection was difficult given the depth and breadth of the responses, a sign the field has more credible teams than early public-goods funding can yet support. That backlog of qualified applicants is part of why the $8 million tutoring grant matters as the program's next, larger tranche.
Why open licensing is the whole point
The licensing requirements are the strategic core, not a compliance footnote. By mandating Apache 2.0 for models and code and Creative Commons Attribution 4.0 for content, Digital Promise guarantees that districts, researchers, and vendors can inspect, adapt, and deploy the tutor without a per-seat license or a vendor lock-in clause. That is a direct answer to the concern that classroom AI is being built behind closed weights the public cannot audit.
Jean-Claude Brizard, CEO of Digital Promise, framed the stakes plainly, warning that without the right foundation, AI will become another barrier to educational progress. Rebecca Griffiths, the program's design director, said educators want AI that understands context and empowers teachers. The open-source mandate operationalizes those views: shared weights and benchmarks let the teaching profession, not just a handful of model providers, shape how the technology behaves in front of students.
The counterweight to closed classroom AI
The timing is pointed. OpenAI, Anthropic, Google, and Khan Academy have all pushed proprietary tutors and teacher assistants into schools this year, each running on closed models with terms the schools accept rather than set. Digital Promise is funding an alternative that districts can host, evaluate, and modify on their own terms. It is a public-infrastructure play positioned against a market that is otherwise consolidating around a few closed vendors.
The math-tutoring focus is a smart place to start. Math has clear right answers, well-understood misconception patterns, and decades of learning-science research, which makes it tractable for building trustworthy benchmarks and evaluation harnesses. Getting an open, well-evaluated math tutor into the wild would give schools a reference point to measure commercial products against, raising the floor for accuracy claims across the category. It also constrains the failure modes that make classroom AI risky: a math tutor that shows its work can be checked against a known answer, which is far harder to do with open-ended writing feedback where the notion of a correct response is contested and hallucinations are easy to miss.
What it means for buyers and builders
For district technology leaders, an openly licensed tutor with published model cards and evaluation tools is a governance gift. It provides a baseline they can host inside their own data boundaries and a yardstick for interrogating vendor claims about accuracy and safety. Even districts that ultimately buy a commercial product benefit, because open benchmarks give procurement teams something concrete to test against rather than trusting a supplier's internal numbers. That shifts leverage toward the buyer in a market where schools have often had to accept accuracy and safety assertions on faith, with no independent way to reproduce the results a vendor cites in a sales deck.
For edtech builders, the program reshapes the competitive terrain. When core model weights, datasets, and benchmarks become public goods, differentiation moves up the stack to integration, classroom workflow, teacher tooling, and outcomes evidence. Vendors that were counting on a proprietary tutor as their moat will need a better one. The companies best positioned are those that can wrap shared infrastructure in a product teachers actually adopt and that can prove learning gains.
The open questions
Execution risk is real. Building an open model that is accurate enough to tutor children in math, safe enough for classroom use, and maintainable after the grant ends is a tall order, and open-source projects can stall once initial funding runs out. The 30 to 36 month window and the reference-implementation requirement are meant to force something durable, but sustainability past the grant is the question that will decide whether this becomes lasting infrastructure or a well-documented demo.
There is also the adoption gap. Public goods only matter if institutions pick them up, and districts short on engineering capacity may still default to a turnkey commercial tutor over a self-hosted open model. The likeliest outcome is a hybrid market where the open weights and benchmarks set the accuracy bar and inform regulation, while commercial vendors compete on the experience layer above them. Technology leaders should watch which grantees ship and how their benchmarks perform.
