Who Is Building AI for African Languages, Including Kiswahili?

TL;DR:

• AI for African languages is being built by the Masakhane research community, South Africa's Lelapa AI, the Lacuna Fund, Microsoft Research Africa in Nairobi, Google Research, and Kenyan teams such as Jacaranda Health.
• Kiswahili is better served than most African languages. Gemini lists it as supported, Anthropic and OpenAI publish Kiswahili test scores, and Google's AI Mode works in Kiswahili. Microsoft 365 Copilot does not list Kiswahili for text prompts (as of November 2026).
• Tests show a real gap. In a 2024 benchmark, GPT-4o scored 87.4 in English and 77.4 in Kiswahili on the same questions. Newer models score higher in Kiswahili, but still below English.
• Sheng and mixed Kiswahili-English messages are harder still.
• Before using AI in Kiswahili with customers, test it on your own real messages and have a fluent speaker check every output.

Who this guide is for

This guide is for Kenyan managers, customer service leads, marketers and NGO staff who are thinking about using AI in Kiswahili. It is also for anyone curious about who is building AI for African languages, and how good it really is.

You do not need a technical background.

The short answer

African language AI is being built by research communities, start-ups, funders, big tech labs and local organisations. Most of the work is in three areas: datasets (examples of a language that AI can learn from), tests that measure how well AI works in each language, and speech tools for people who would rather talk than type.

Kiswahili gets more attention than most African languages. Smaller Kenyan languages such as Dholuo, Kalenjin and Kikuyu have far less support, though that is starting to change.

Masakhane: Africans building language AI together

Masakhane is a grassroots organisation that aims to strengthen language AI research in African languages, "for Africans, by Africans". The name roughly means "we build together" in isiZulu. Its website reports more than 1,000 participants from 30 African countries, a figure it dates to February 2020.

Its projects include MasakhaNER, a dataset that teaches AI to recognise names of people, places and organisations in African languages. The first version (2021) covered ten languages; MasakhaNER 2.0 (2022) covered 20.

In January 2026 the Masakhane African Languages Hub invited proposals to build datasets for 50 African languages, with funding from Google.org, the UK's FCDO, Canada's IDRC and the Gates Foundation, as HapaKenya reported.

The Hub also runs LINGUA Africa, an open call, with Microsoft's AI for Good Lab, the Gates Foundation and Google.org. Microsoft's programme page lists 26 selected projects covering more than 50 African languages, dialects and sign languages, spoken across 47 countries. Five involve Kiswahili, and three of those are based in Kenya: AFYA-LINGUA (World Health Organization), RADA (Algedi Intelligence Labs) and SemaShambani (AntuGrow). These are early-stage projects, so watch them rather than expect finished tools today.

Lelapa AI: a start-up building for African languages

Lelapa AI is a South African start-up. Its website says it focuses on contact centres in telecoms and financial services. Its team page lists Pelonomi Moiloa as CEO.

InkubaLM is Lelapa's small open language model. Its model card on Hugging Face says it has about 0.4 billion parameters (a measure of size) and covers isiZulu, Yoruba, Swahili, isiXhosa and Hausa, plus English and French. The big models behind everyday chatbots are far larger.

The model card is open about the limits. Outputs may be inaccurate or biased, the model may switch languages mid-answer, and users should test it for safety before building anything with it. The licence (CC BY-NC 4.0) allows non-commercial use only. So for a Kenyan company, InkubaLM is a research building block. If your team wants to try open models on its own machines, see our guide on running open-source AI models on your own computer.

Vulavula is Lelapa's product for transcribing calls and translating text. Its language support page, checked in November 2026, marks isiZulu and South African English transcription as production-ready. Swahili translation is in "Beta" (early access), and Swahili transcription is "coming soon".

That is a useful reality check. Even a company built entirely around African languages is still working through Kiswahili, which shows how much work each language needs.

Lacuna Fund: paying for the missing data

Lacuna Fund is a group of funders, including the Rockefeller Foundation, Google.org, IDRC and the Wellcome Trust, that pays for labelled datasets in low- and middle-income countries. It funds data in agriculture, language, health and climate.

Data is the quiet foundation of language AI. A model can only learn a language from examples of it, and for Kenyan languages those examples have been thin.

Its catalogue lists several Kenyan language datasets, mostly from researchers at Maseno University, the University of Nairobi and Africa Nazarene University:

• KenSpeech: transcribed Swahili speech.
• KenTrans: matching texts in Swahili and local Kenyan languages, used to train translation.
• KenSwQuAD: a question-and-answer dataset in Swahili.
• Kencorpus and KenPos: text collections in Kenyan languages for training and analysis.

These are raw material for researchers and builders, not consumer apps.

Microsoft: Project Gecko and Paza, from Nairobi

Microsoft Research Africa in Nairobi is at the centre of Microsoft's Kenyan language work.

Project Gecko was announced in November 2025, HapaKenya reported. Microsoft says it is co-led by three of its research teams, including Microsoft Research Africa in Nairobi. It started with farming in Kenya and India, building on Digital Green's FarmerChat assistant, and used about 3,000 hours of crowd-sourced Kenyan speech.

Paza is the Nairobi team's speech recognition work. Microsoft's labs page describes PazaBench as the first speech recognition benchmark for low-resource languages, testing 52 models across 39 African languages; the current Project Gecko page gives 53 models and 61 languages. Microsoft has released speech models for six Kenyan languages: Swahili, Dholuo, Kalenjin, Kikuyu, Maasai and Somali. In September 2026 Microsoft wrote that the models were tested with farmers on ordinary mobile phones.

Speech matters because many Kenyans would rather speak to a phone than type, especially in their first language.

Google: WAXAL, Translate and AI Mode in Kiswahili

WAXAL. On 6 March 2026 Google Research described WAXAL, an open speech dataset for 27 Sub-Saharan African languages, with about 1,846 hours of transcribed speech and more than 565 hours of recordings for text-to-speech. Partners included Makerere University and the University of Ghana. It is free to reuse under a CC-BY-4.0 licence.

Google Translate. In June 2024 Google added 110 languages to Translate, about a quarter of them African. Luo (Dholuo) was one.

AI Mode in Kiswahili. In March 2026 Citizen Digital reported that Google had brought AI Overviews and AI Mode in Search to 13 African languages, including Kiswahili and Somali, with the first rollout aimed at Kenya, Nigeria, Ethiopia and Tanzania.

Jacaranda Health: a Kenyan team building in Kiswahili

Some of the most useful work happens inside Kenyan organisations that need Kiswahili to serve real people.

Jacaranda Health is a Kenyan non-profit that texts health advice to pregnant women and new mothers through its PROMPTS service. A May 2025 IDRC story describes its UlizaLlama model as "the world's first large language model (LLM) for health in Kiswahili", and says it is open source.

The same story is frank about limits. Jacaranda estimated the system still made errors in 30% of cases, and a human takes over when a mother is not satisfied with the automated answer. Even a team that built its own Kiswahili model kept people in the loop.

How well do the main chatbots handle Kiswahili?

Most big tools read and write standard Kiswahili reasonably well, but quality is lower than in English and varies by tool and task. Here is what each company says, as of November 2026.

The Copilot point surprises many Kenyan teams. Microsoft's page says a prompt in an unsupported language may get a message asking you to change it. If you plan to use Copilot in Kiswahili, test it on your own work account first. Copilot stays very useful for English work in Word, Excel, Outlook and PowerPoint.

A language being "supported" tells you the tool will try. It does not tell you the result is good enough for your customers.

What the published tests say

The tests point the same way: Kiswahili trails English, newer models score higher in Kiswahili than older ones, and smaller African languages are further behind.

IrokoBench (2024)

IrokoBench tests AI in 17 African languages. It was led by David Ifeoluwa Adelani with 26 co-authors. On its general knowledge test, GPT-4o scored 87.4 in English, 77.4 in Swahili, and 60.0 on average across the African languages. The authors report a gap of up to 28% between African languages and English for the best model. Swahili did far better than the African average, but this was a 2024 model.

OpenAI's multilingual results

OpenAI publishes results for an exam-style test translated by professionals into 14 languages. In the file we checked, o3-high scored 0.860 in Swahili and 0.780 in Yoruba. GPT-4o mini scored 0.619 in Swahili. The file covers models up to o3, o4-mini and GPT-4.1, so OpenAI's newest models do not appear.

Anthropic's results for Claude

Anthropic reports each Claude model's score in each language as a percentage of the same model's English score. For Swahili: Claude Fable 5.2, 99.2%; Claude Opus 5.5, 98.3%; Claude Sonnet 5.5, 96.9%; Claude Haiku 5.5, 94.0%. Yoruba was lower for every model. Anthropic says to compare languages within a model, not to rank models.

The lesson for buyers

Across both companies, the smaller, cheaper models lose more quality in Kiswahili. The model running inside a product may be weaker than the flagship you saw in a demo, so ask your vendor which one it uses.

All these tests use exam-style questions in standard written Kiswahili. Your customers write short WhatsApp messages, mix in English, use Sheng, and mention M-PESA and matatus. We did not find a published Kiswahili score for Gemini on Google's own pages, so you will need to test it yourself, as with every tool.

Sheng and code-switching: the harder test

Sheng and everyday mixed Kiswahili-English are where AI struggles most.

The clearest study we found is "Beyond Metrics" by Millicent Ochieng of Microsoft Research Africa and colleagues, published at the AfricaNLP workshop at ACL 2025. They tested seven AI models on WhatsApp chats in Swahili, English and Sheng from young people in Nairobi health support groups. Some models scored well on the numbers but still missed meaning and context. GPT-4 and GPT-4-Turbo handled the mixed language and context well and agreed most with human judges. Even so, the models struggled with cultural nuance, especially outside English.

A 2025 paper by Kezia Oketch, John Lalor and Ahmed Abbasi collected 2,170 free-text answers from Kenyan speakers. Their language mixed tribal influences, urban slang, code-mixing and loanwords, and that variety affected how well models performed.

Both studies used models that are now a year or two old. Newer tools may do better. The practical point holds: if your customers write in Sheng, test the tool on Sheng.

What a Kenyan organisation should do now

Treat Kiswahili AI output as a first draft until your own tests prove otherwise. Most teams can run this in a week.

Step 1: Collect 30 real messages

Pull 30 real customer messages: 10 in standard Kiswahili, 10 in mixed Kiswahili and English, and 10 in Sheng or very informal language. Remove names, phone numbers and account details first. Kenya's Data Protection Act applies to any personal data you put into an AI tool; our guide on whether staff can use ChatGPT under the Data Protection Act explains what to watch.

A typical mixed message: "Hi, nimetuma pesa jana lakini sijapata confirmation. Mnaweza check?"

Step 2: Test two or three tools on the same messages

Give each tool the same messages and instructions, starting with the free versions. Use the four Cs: Clarity, Context, Constraints and Customisation. A prompt you can copy:

You are helping a customer service team at a Kenyan [type of business]. Below is a customer message. First, tell me in English what the customer wants and how they feel. Then write a polite reply in simple Kiswahili, under 60 words. Use everyday words a Nairobi customer would use, not textbook Kiswahili. Do not promise refunds or dates. If anything is unclear, say so instead of guessing.

Customer message: [paste message]

For notices and translations:

Translate this English notice into simple Kiswahili for parents in Kenya. Keep it short and warm. Then list any words or phrases you were unsure about, and why.

Notice: [paste notice]

Asking the tool to flag its doubts will not catch every error, but it often points your reviewer to the weak spots.

Step 3: Have fluent speakers score each reply

Ask two fluent Kiswahili speakers, ideally one comfortable with Sheng, to score each reply from 1 to 3:

Step 4: Decide where AI helps

Step 5: Write the rule down

Put the decision in your AI use policy: which tools are approved for Kiswahili, for which tasks, and who checks the output. Test again every few months, because the tools change fast. Our guide to training customer service teams on AI covers the wider set-up.

What goes wrong, and how to avoid it

Trusting a demo. A vendor shows a perfect reply to a simple question. Your customers ask messier ones. Test on your own messages.

Textbook Kiswahili. Many tools write stiff, formal Kiswahili that sounds odd on WhatsApp. Ask for simple, everyday language and show an example reply your team likes.

Missed local meaning. AI can confuse similar words or miss what a phrase means in Kenya. A fluent reviewer catches this; a spell-checker does not.

Forgetting other Kenyan languages. If your customers prefer Dholuo, Kikuyu or Kalenjin, mainstream chatbots will usually be much weaker. Test any tool carefully first.

No human on sensitive messages. Health, money, legal and safeguarding messages need a person to read them first.

Marketing teams writing Kiswahili posts can find more on tone and accuracy in our guide on how marketing teams in Kenya use AI well and legally.

Questions people ask

Can ChatGPT write in Kiswahili?

Yes. ChatGPT reads and writes Kiswahili, and OpenAI publishes Kiswahili test results. Quality is lower than in English, and smaller models are weaker. Have a fluent speaker check anything customers will see.

Does Microsoft Copilot support Kiswahili?

Not for text prompts, according to Microsoft's support page updated in September 2026. Swahili is listed only for Copilot Voice on personal accounts. Test it on your own work account before relying on it.

Which AI tool is best for Kiswahili?

There is no fixed answer, and the tools change every few months. Published tests suggest the largest models from OpenAI, Anthropic and Google handle standard Kiswahili well. Test two or three on your own messages and let a fluent speaker judge.

Can AI understand Sheng?

Partly. Microsoft Research Africa found that models struggled with cultural nuance in Sheng and mixed-language WhatsApp chats, even when their scores looked good. Test on real Sheng messages first.

Is there an AI model built for Kiswahili?

A few. Jacaranda Health's UlizaLlama is an open model for health questions in Kiswahili, and Lelapa AI's InkubaLM includes Swahili among five African languages. For most businesses they are building blocks rather than ready-made tools.

What about Kikuyu, Dholuo and other Kenyan languages?

Support is much thinner. Microsoft has released speech models for Swahili, Dholuo, Kalenjin, Kikuyu, Maasai and Somali, and Google Translate added Luo in 2024. Mainstream chatbots are generally weaker in these languages.

Where FuKazee fits in

FuKazee trains Kenyan teams to use AI for the work in front of them, including work in Kiswahili. We assume nobody in the room knows AI yet, avoid jargon, and show the free version of every tool so people can keep practising. We never push one tool. We teach the four Cs, safe handling of customer data, and how to check every output, which matters even more in Kiswahili. Since 2022 we have trained more than 3,000 people.

For customer service, marketing and communications teams, we can build a session around your own messages, using the testing process in this guide. See our page on AI training for customer service teams in Kenya, or AI training in Africa for teams across the continent.

To discuss a session on using AI in Kiswahili safely, email us at lead@fukazee.com

Sources

All sources last checked 11 November 2026.

• Masakhane, official website
• Adelani et al., "MasakhaNER: Named Entity Recognition for African Languages" (arXiv, 2021)
• Adelani et al., "MasakhaNER 2.0: Africa-centric Transfer Learning for Named Entity Recognition" (arXiv, 2022)
• HapaKenya, "Masakhane Hub launches landmark funding initiative to scale AI for 50 African languages" (12 January 2026)
• Microsoft Research, LINGUA Africa programme page and awardees
• Lelapa AI, official website
• Lelapa AI, team page
• ITWeb, "SA AI start-up in push to meet diverse linguistic needs" (21 October 2024)
• Lelapa AI, InkubaLM-0.4B model card on Hugging Face
• Lelapa AI, Vulavula language support
• Lacuna Fund, official website
• Lacuna Fund, datasets catalogue
• HapaKenya, "Microsoft launches AI to empower farmers with speech models in Kikuyu, Swahili and Maa" (18 November 2025)
• Microsoft Research, Project Gecko
• Microsoft Azure AI Foundry Labs, Paza
• Microsoft Source EMEA, "Africa needs AI that serves its priorities" (17 September 2026)
• Google Research, "WAXAL: A large-scale open resource for African language speech technology" (6 March 2026)
• Google, "110 new languages are coming to Google Translate" (June 2024)
• Citizen Digital, "Swahili, Somali among 13 African languages added to Google AI Search" (6 March 2026)
• Google, "Where you can use the Gemini web app"
• IDRC, Jacaranda Health and UlizaLlama story (20 May 2025)
• Microsoft Support, "Supported languages for Microsoft Copilot" (updated September 2026)
• Adelani et al., "IrokoBench: A New Benchmark for African Languages in the Age of Large Language Models" (arXiv)
• OpenAI, Multilingual MMLU benchmark results (GitHub)
• Anthropic, Claude multilingual support documentation
• Ochieng et al., "Beyond Metrics: Evaluating LLMs' Effectiveness in Culturally Nuanced, Low-Resource Real-World Scenarios" (Microsoft Research)
• Oketch, Lalor and Abbasi, "Benchmarking Sociolinguistic Diversity in Swahili NLP: A Taxonomy-Guided Approach" (arXiv, 2025)