What LLM Support for Swahili Actually Looks Like Today

Every current frontier model will answer you in Swahili. That tells you almost nothing, because “support” is at least five separate properties and a model can be strong on the first two and weak on the rest. Here is how to check each one, and which public datasets to check it against.

Five things “support” can mean

• Tokenisation efficiency. How many tokens a Swahili sentence costs relative to its English equivalent. This is a property of the tokeniser, is fixed for a given model, is measurable exactly, and it sets your price and your effective context length. Swahili is agglutinative — nitakupenda is ni-ta-ku-pend-a, four morphemes and a final vowel in one orthographic word — so a tokeniser trained mostly on English tends to split it into several pieces. See the token cost of Swahili.
• Generation fluency. Whether the output is grammatical and idiomatic. This is where noun-class concord lives and it is the property most worth probing, because errors here are systematic rather than random.
• Instruction following in-language. Whether a Swahili prompt produces the same behaviour as its English translation, or whether the model drifts into English, ignores format constraints, or follows the instruction less exactly. Frequently the largest gap of the five and the least reported.
• Knowledge and factual grounding. Whether the model knows about Tanzanian and Kenyan subject matter, or knows global subject matter and merely renders it in Swahili. These are different capabilities and the second is much more common.
• Safety and moderation parity. Whether refusals, filters and jailbreak resistance behave the same as in English. Safety training is heavily English-weighted, and the asymmetry is a known class of problem covered in why low-resource languages hallucinate more.

Where Swahili sits, and what is public

Swahili is among the best-resourced sub-Saharan African languages, and that is a statement about the field rather than an abundance. It is an official language of Tanzania and Kenya, is widely used across the East African Community, and was adopted as an official working language of the African Union in 2022. Commonly cited speaker totals range from roughly 80 million to over 200 million depending on whether second-language speakers are counted — the spread is real and is about methodology, not uncertainty, so quote a range rather than a number.

What you can check, all published and dated by their publishers:

• FLORES-200, published by Meta — a parallel evaluation set covering 200 languages, including Swahili as swh_Latn. It is the standard reference for translation quality across low-resource languages and its per-language scores for the NLLB-200 models are published with it.
• Masakhane — the community research effort behind a family of African-language datasets (named-entity recognition, news classification, machine translation) built with native speakers. If you need Swahili evaluation data that was not machine-generated, this is where to look first.
• Common Crawl’s published crawl statistics — per-language page counts for the web corpus most training sets are derived from. This is the honest way to answer “how much Swahili is in the training data”: no frontier lab publishes its own mixture, so any specific percentage attributed to a commercial model is somebody’s estimate. Check the crawl statistics yourself and treat the result as an upper bound on availability, not as a training share.

The probe that matters: noun-class concord

Swahili has a noun-class system — conventionally around fifteen to eighteen classes in pairs of singular and plural — and class membership controls agreement on adjectives, numerals, demonstratives, possessives and the verb. Change the noun and a chain of prefixes changes with it:

class 7/8   ki- / vi-
  kitabu kimoja kizuri kilianguka      one good book fell
  vitabu viwili vizuri vilianguka      two good books fell

class 1/2   m- / wa-   (people)
  mtoto mmoja mzuri alianguka          one good child fell
  watoto wawili wazuri walianguka      two good children fell

class 3/4   m- / mi-   (trees, natural things)
  mti mmoja mzuri ulianguka            one good tree fell
  miti miwili mizuri ilianguka         two good trees fell

class 5/6   ji-/Ø / ma-
  tunda moja zuri lilianguka           one good fruit fell
  matunda mawili mazuri yalianguka     two good fruits fell

Four prefixes change per sentence, and none of them is predictable from the English. This is the ideal probe because it is unambiguous — there is one correct answer, it does not depend on register or dialect, and a model that has genuinely learned Swahili morphology gets the whole chain right while a model that has learned Swahili vocabulary gets the noun right and the concord wrong. It also degrades gracefully as a measure: you can count how many of the four slots are correct rather than scoring the sentence pass or fail.

A probe set you can run

Six probes, each targeting one of the five properties, each with a checkable answer. Run them yourself; do not take anyone’s word for the result, including this page’s.

1. Concord chain. “Translate into Swahili: three large books fell from the table.” Check the numeral, adjective and verb prefixes against the class of the noun the model chose. Repeat with children, trees and fruit to cover four classes.
2. Agglutinative verb construction. Ask for the Swahili for “I will not have seen you” and for the morpheme breakdown. A model that can segment sitakuwa nimekuona into its parts is doing morphology; one that produces the string but cannot segment it is retrieving a phrase.
3. Instruction following in-language. Take a prompt you already trust in English — with a format constraint, such as “reply with exactly three bullet points and no preamble” — translate it into Swahili, and compare compliance, not quality. Drift back into English is itself a result.
4. Local knowledge. Ask about something specifically East African rather than globally famous, where you can verify the answer. A model that answers in fluent Swahili and hallucinates the content has failed a different test from a model that answers in broken Swahili.
5. Register and Sheng. Ask for the same message in Kiswahili sanifu (standard Swahili) and in the Nairobi urban vernacular. Models often collapse the distinction, which matters if your users write to you in the second one.
6. Safety parity. Take a prompt your English pipeline correctly refuses, and its Swahili translation. If the two behave differently, that difference is a production issue regardless of which direction it goes.

Reading the results

Score the probes separately and resist collapsing them into one verdict. The common shape of the result is a model that is fluent and confident, mostly right on concord for the two commonest noun classes and weaker on the rest, noticeably less exact at following formatting instructions in Swahili than in English, and thinner on East African specifics than on global ones. Which of those matters depends entirely on what you are building: a summariser cares about fluency, a structured-extraction pipeline cares about instruction following, and a support router cares about the vernacular probe.

One thing worth checking on the same day is whether a smaller, cheaper model is materially worse for your specific task, because the gap between models tends to be larger on lower-resource languages than on English — so a routing policy tuned on English traffic can be badly wrong here.