A Public Bet on Open Tutoring
Digital Promise has opened a request for proposals that stands out in an edtech market saturated with proprietary AI. Through its K-12 AI Infrastructure Program and backed by the Gates Foundation, the organization is offering up to 8 million dollars to a single team to build an open-source, education-specific AI model for tutoring. The program is called Open Source AI Model for Tutoring, and its premise is a quiet rebuke to the dominant approach: rather than wrap a general-purpose frontier model in an education skin, build a model designed for learning from the ground up and release it openly.
The framing matters because most AI tutors on the market today are closed systems built atop commercial models that were never optimized for pedagogy. Digital Promise is arguing that the infrastructure students learn on should be a public good, inspectable and improvable by the education research community rather than a black box controlled by a single vendor. In a sector where trust, transparency and evidence matter more than in most, that is a substantive position, and eight million dollars behind it is a meaningful signal.
Fixing What Current Tutors Get Wrong
The RFP is refreshingly specific about the problems it wants solved. Current tutoring models, it notes, tend to provide answers too quickly, are excessively verbose and have weak awareness of student knowledge and motivation. Anyone who has watched a general chatbot tutor a student recognizes these failure modes. A good human tutor withholds the answer, prompts the student toward it, reads confusion and adjusts. General-purpose models, tuned to be helpful and complete, do the opposite by defaulting to fast, thorough answers that short-circuit the learning process.
This is where an education-specific model earns its keep. Building a tutor that applies learning science principles, that knows when to hold back the answer and how to gauge what a student actually understands, requires training choices and evaluation methods that a general model does not make. Bryan Richardson, describing the goal, said the RFP seeks teams focused on developing the best AI tutoring model using cutting-edge methods and applying learning science principles. That pairing, frontier technique with pedagogical grounding, is the harder and more valuable target.
Open by Design
The openness requirements are strict and deliberate. The winning team must deliver model weights, training and fine-tuning code, datasets, evaluation tools, documentation and reference implementations, all under open licenses. Content is to be released under Creative Commons Attribution 4.0, and software and code under Apache 2.0. This is not open in the marketing sense of a public API. It is open in the sense that the entire artifact, from weights to evaluation harness, becomes shared infrastructure that any district, researcher or vendor can build on and audit.
That structure changes the incentives. A single funded team produces the foundation, and the whole ecosystem inherits it, avoiding the duplication of every edtech company training its own opaque tutor. It also makes the model accountable to scrutiny in a way closed systems are not. Educators and researchers can examine how the tutor was trained, test it against their own evidence standards, and improve it. For a technology being placed in front of children, that inspectability is not a nice-to-have. It is close to a prerequisite for legitimate trust.
Serious Requirements for Serious Work
Digital Promise is not casting a wide net. The eligibility bar is high, requiring prior experience with large language models, at least one peer-reviewed publication before May 8, a record of contributing digital public goods and meaningful prior deployment using real student or user data. Teams must combine machine learning engineering, K-12 classroom practice, learning science and edtech product partnerships, and at least one major tutoring edtech provider must be identified or conditionally committed. This is a demand for demonstrated, interdisciplinary competence, not aspiration.
The timeline reflects the ambition. Applications are due July 31, and the grant period runs 30 to 36 months with work beginning in November. That multi-year horizon is a signal that Digital Promise understands what it is asking for. Building a genuinely effective, well-evaluated tutoring model is a research program, not a product sprint. The requirement that a real edtech provider be part of the team is the pragmatic touch, ensuring the resulting model has a path into actual classrooms rather than ending as an academic artifact.
Why It Matters for the Sector
The generative AI edtech market is growing fast, and most of that growth is proprietary. An 8 million dollar public investment in open tutoring infrastructure is small against that tide, but its influence could be outsized. If the resulting model becomes a credible open baseline, it gives schools an alternative to locking their students' learning experiences into a single vendor's closed system, and it gives smaller edtech companies a foundation they could never afford to train themselves. Public infrastructure has a way of reshaping markets that pure competition does not.
There is also a values argument that CIOs and education leaders should weigh. Buyers in this sector consistently say they want teacher control, audit trails, privacy protection and clear data rules before they trust AI tutors. An open, education-specific model addresses several of those demands directly by making the system inspectable and improvable. Whether this particular grant produces a model that meets its ambitions is unknowable today. But the direction, treating the AI that teaches children as public infrastructure rather than a proprietary black box, is one the sector needs.
