Bantu Capability Index

Measuring what AI models can actually do in Bantu languages, one capability at a time.

Returning user
Already have an account? Jump back in.
Login →
New here?
Create an account to follow every component as it is scored.
Create an Account

One index, many components

Each component measures one capability, has its own board and its own status. Families group them; new components join as their ground truth matures.

5families
3components today
1with a published board
459released syllable inventories (FSIs): the width every component grows toward
Foundations

the building blocks a language is made of

L26
Operating AlphabetOperational

Can a model list every building block its language is made of? English has 26 letters; Bemba has 480 syllables.

  • 6 of 459 languages scored
  • Live model board at l26.ai
  • board v1.0
Open l26.ai →
Ground truth in this family: Syllable inventories (459 languages), feeds L26 · Tone & homographs (6 languages)
Number system

counting, and everything built on it

N12
NumeracyWorking prototype

Can a model operate the number system, or only recall number words? Money, dates and sums are TRACKS inside it, not separate components.

  • 10 of 459 languages scored · 20 admitted
  • Full board, S10: Claude Opus 5.5 52.6 · Claude Sonnet 5.5 43.3 · Claude Haiku 4.5 17.3 (of 100)
  • N12 v1.0
Open the N12 board →
Ground truth in this family: Bantu Calculator (19 languages), feeds N12
Grammar

the machinery that makes a sentence agree with itself; every component here is a dimension group of the concord matrix

—
No component yetPlanned

The family is declared; a scored component joins once its ground truth is ready.

Ground truth in this family: Nouns (352 languages) · Verbs (6 languages) · Concord (20 languages)
Language in use

what people actually say to each other

—
No component yetPlanned

The family is declared; a scored component joins once its ground truth is ready.

Ground truth in this family: Stories (2 languages)
Applied

capability where it has consequences

H200
Clinical languageWorking prototype

Can a model name the body, understand what a patient says, and say it back? And what does it invent when it cannot — on danger signs?

  • 24 of 459 languages scored
  • Public index not yet published: the answer key is in its review pass first.
  • H200 v0.1-prototype
No public page yet.
Ground truth in this family: Body & Health (24 languages), feeds H200

How it works

The same principles hold for every component.

Native ground truth

Every answer key comes from native speakers, and every attested alternative form is accepted.

One capability each

A component measures a single capability, such as the alphabet, numbers or clinical language, so a score says exactly what was tested.

Keys stay private

Answer keys are never published. Models are scored blind, and editions are frozen so results stay comparable.

Width, depth, scope

Components grow across more languages, more items per language, and more of the standard, measured against 459 released inventories.