Imagine an AI that can solve a math problem in a second. A student writes 1/2 + 1/3 = 2/5, getting it wrong, and the AI replies only, "That's wrong, the answer is 5/6." Its arithmetic may be flawless, but it isn't a good teacher. Knowing the answer and helping a student arrive at it on their own are two different abilities. OmniEdu is an attempt to put both into a single model.
OmniEdu is an open foundation model aimed at K-12 education. A foundation model is a "base" model trained broadly so it can be applied across many tasks. OmniEdu takes that base and layers education data on top to raise it into a specialist. It comes in three sizes — 4B, 9B, and 27B. The "B" means the parameters (the values a model adjusts through training) number in the billions. Parameters are closer to the "brain cells" that hold a model's knowledge and judgment — the more there are, the subtler the patterns the model can learn, but the more compute and cost it takes.
One decision sits at the core of this model: the training data was reorganized around "what ability does it build (competency)" rather than "where did it come from (source)." Why that matters becomes tangible if you follow that one wrong fraction all the way through.
1/2 + 1/3 = 2/5. The student added the numerators together and the denominators together. A common mistake. Faced with this one line, a good teacher does four things at once.
First, they know the right answer. It's 5/6. This is subject competency — the skill of actually solving the problem correctly.
Next, they know where this problem sits. They recognize that it's addition of fractions with unlike denominators — in other words, a unit on finding a common denominator. Knowing which junction of the curriculum the current material falls on is curriculum grounding.
Then they diagnose why it went wrong. Not "the student can't calculate," but the specific misconception: "the denominators were different, yet the student just added straight across." This is diagnostic reasoning. If you don't know where the slip happened, the next explanation goes nowhere.
The last one is the hardest. They don't hand over the answer. Instead they ask back, "The denominators are different — can we really add them as they are?" They lay down a single foothold so the student recalls common denominators on their own. This is pedagogical action and scaffolding.
| Competency | What it does in front of the wrong fraction |
|---|---|
| Subject competency | Knows the answer is 5/6 |
| Curriculum grounding | Identifies it as a "common denominator" unit problem |
| Diagnostic reasoning | Pinpoints the misconception of adding denominators straight across |
| Scaffolding | Gives a hint — "can we add them as they are?" — instead of the answer |
A model that is only good at solving problems stops at the first cell. A tutor model that is only good at conversation can mimic the fourth cell but is shaky on the first. What OmniEdu set out to do was fill all four cells inside one model.
The problem is that, until now, education AI data wasn't collected in a way that evenly covered these four cells.
The old approach grouped training data by source and by task. It's similar to a library organized only by publisher. The shelves look tidy, but the view of "what ability does this book build" is missing entirely. As a result, the abilities that are unique to education — diagnosing, scaffolding — end up sparsely represented in the data.
So the researchers changed the organizing criterion itself to competency. They took more than 100 educational resources and general instruction datasets and rearranged them along the four axes above.
The important part is that they didn't just scrape up as much as possible. The researchers first used mechanical rules to strip out obvious errors and duplicates, then went through the remaining content meaning by meaning to check whether it actually made sense and rewrote it where needed, and scored each piece for quality against criteria suited to its purpose, keeping only the strong examples. On top of that, within a set budget of tokens (the smallest unit a model uses to process text), they selected for diversity so that similar types wouldn't just pile up, and attached to each example an instruction spelling out "this should be answered with such-and-such pedagogical intent." It was carefully refined across several stages.
The result of this curation was 69,999 examples, of which 60,951 are education-specific. In terms of volume, that's roughly 15.96 million supervised response tokens. Put plainly, around 100 sources were filtered and refined over multiple passes down to 70,000. Using this data, the researchers fine-tuned (the process of adding specialized data to a general model to make it a specialist) the three model sizes.
Evaluation ran along three axes — curriculum grounding, problem solving, and tutoring — plus general ability. The results for the largest model, OmniEdu-27B, are below. All percentages can be read against a maximum of 100.
| Benchmark | What it measures | 27B score |
|---|---|---|
| MathFish | Judging whether a problem matches curriculum standards | 85.89% |
| EDUMATH | Generating problems tailored to the curriculum and student interests | 86.95% |
| K12-Bench | Curriculum understanding — grasping concepts, prerequisites, and position within the curriculum (EM = share of answers that exactly match, F1 = a score combining precision and recall) | EM 63.12% / F1 76.69% |
| MathTutorBench | Tutoring that guides with scaffolding (hints) | 78.74% |
| LongTutor | Quality of long-form tutoring | 3.02 |
One thing first. The 3.02 at the bottom, for LongTutor, is not a percentage but a score rating tutoring quality. Its scale differs from the % metrics above, so you shouldn't line it up with them and say something like "it's the lowest of the three." It's a number that has to be read on its own.
Taken together, OmniEdu scored evenly well on both the ability to align with and understand the curriculum (MathFish, EDUMATH, K12-Bench) and the ability to guide students (MathTutorBench, LongTutor). Problem-solving ability itself was evaluated separately, outside the table above, with GAOKAO-Bench, EXAMS-V, and MDK12-Bench. In other words, there's no visible trade-off where lifting one ability comes at the cost of another. That suggests the strategy of rebuilding the data around competencies paid off.
That said, let's be clear about the limits of interpretation. The abstract only states that these scores "went up"; it gives almost no direct comparison figures, such as how many points higher it is than some other model. The only comparison stated explicitly is that OmniEdu had the highest average tutoring quality on LongTutor among the models evaluated — and it's notable that this metric happens to be the one measuring the "ability to teach." So the numbers in the table above should be read as "roughly where OmniEdu landed," not as wins and losses against other models — reading them that way goes beyond what the evidence supports.
Judging from the abstract alone, a lot is still open.
If you're considering adoption, these four need to be checked separately against the original paper and follow-up evaluations.
No. Diagnosing misconceptions or offering scaffolding is closer to an assistive tool that supports a teacher's "work." The final call on which scaffold to lay down, and the human connection with the student, still belong to the teacher.
The abstract states that the benefits of education-oriented tuning appeared consistently across all three sizes. That means even a small model still sees the gains of competency tuning. The absolute scores are higher for the larger models, though, so choose based on whether you prioritize quality or cost.
ONTLE is a service that dispatches an AI project manager to the side of small businesses and solo founders worn down by repetitive work. What's worth taking from OmniEdu isn't the model — it's one piece of its design philosophy: the idea of treating "giving the answer for someone" and "helping them judge for themselves" as different abilities from the start, and splitting the data accordingly.
Suppose a client sends the message, "Please push next week's scheduled delivery back by two weeks." The simplest AI just changes the date on the schedule. In terms of the fraction earlier, it stopped at the first cell of "knowing the answer." An AI PM that reaches the fourth cell moves differently. It sees where this delivery sits along the project's path (context), identifies whether the real bottleneck is materials or staffing (diagnosis), and instead of ending with a one-liner "done, pushed it back," it lays out the downstream schedule that now slips and the available options, handing back to the owner a decision to make rather than a fait accompli (scaffolding).
One thing must be stated plainly, though. What OmniEdu showed in numbers is education benchmark scores, not performance in a real workplace. So "just take this model and use it as is" doesn't hold. ONTLE's AI PM, too, can become a more useful aide when its data is gathered along competency axes — "read the context," "diagnose where things are stuck," "organize the next decision instead of spoon-feeding it" — rather than along a task list like "write reports" and "coordinate schedules." This isn't an already-validated application; it's a hypothesis we have to verify for ourselves.