Resource
Fab AI benchmark leaderboard
Fab AI evaluates models on education tasks that matter where budgets, bandwidth and language coverage are constrained in the low- and middle-income countries the benchmark is built for, and publishes the results as a sortable leaderboard rather than a paper. The core table scores a pedagogy benchmark alongside SEND, visual maths and visual reasoning, with separate tabs for multilingual pedagogy, foundational literacy and read-aloud assessment.
What you can do with it
Compare roughly 300 models on pedagogy scores that currently run from the low nineties down to 21%, then sort by what an education ministry actually has to weigh: price per million tokens, parameter count, open weights and release date. The spread at the bottom is the useful part. Sub-2B open models sit between 21% and 43% on the pedagogy benchmark while frontier models sit in the low nineties, which turns "can we run this offline on cheap hardware" into a quantified trade rather than a guess.
The multilingual pedagogy tab is the reason to bookmark the page. Most education benchmarks are English-only, and the question of whether a model's pedagogy survives translation is exactly the question a ministry in a multilingual system needs answered. The data is downloadable as CSV, so the comparisons can be re-run locally.
Who it is for
Policymakers and procurement teams choosing models for national systems, researchers studying the access gap in AI-supported education, and developers who need a pedagogy target that is not an English-language exam.
Notes and caveats
A leaderboard measures models on the benchmark's own tasks: a high pedagogy score is evidence about those items, not about a classroom. The price column blends input and output token prices at a 3:1 ratio, which is a modelling choice and not what any particular deployment will cost. Coverage of languages and of locally hosted models moves as new models appear, and the low-end scores for small models partly reflect task format as much as capability. Fab AI's framing is explicitly about low- and middle-income countries, so the benchmark's item set reflects that context and should not be read as a general education ranking. Check release dates: several entries predate the current model generation.