Build African-language AI on consented data.
Member
I recorded directions from the market in Yoruba.
Nzinga App
Who can use this recording? Only me, Nzinga's models, Nzinga and approved partners, or Everyone.
Member
Nzinga's models, please.
✓ Choice saved · 2:05
Works with, measured on and released through
From speech to a trained model.
Five building blocks, one protocol. Each recording keeps its speaker's choice from the phone to the model.
See consent in action on your phone.
Install Expo Go first
From your phone's app store, free to install.
Scan the code
Or open the exp link.
Enter code 482917
Any phone number works, then code 482917 to sign in.
Use any invite
Any eight characters, then record your first note.
Measured, not promised.
Tokens per language on FLORES+, 1,012 sentences each. Method in the docs.
“Hausa needs 1.43 tokens for every English token with Nzinga, against 1.55 with the best other.”
Every speaker in a recording has their own consent record, never edited.
Speakers decide how recordings travel.
Consent for every recording, contacts hashed and never kept, and audio on the phone today. Honest about what is not built.
Consent
Each person in a recording chooses for their own voice.
Contacts
Hashed on the phone; the server keeps none of the codes.
Sign-in
Codes kept as hashes, expiring after ten minutes.
Withdrawal
Withdraw any recording at any time. You keep your pay.
Members share in every licence fee.
Members are paid per accepted minute, at a rate set at launch, with a licence share on top.
Rates are set per market at launch. Members keep what they are paid.
Per accepted minute, with a licence share and three invites.
Licensed, consented recordings in African languages, every speaker's choice checked.
Recordings shared as Everyone, released as open data.
What shipped most recently.
v2 benchmark results
Fewer tokens than twelve other tokenizers on all 12 Tier 1 languages, 22% fewer than the best.
Join, record, library and wallet
The phone app runs in demo mode in Expo Go, with every Nzinga screen.
Sign-in, invites, consent, earnings
The pilot API server, tested end to end with 21 tests.
Twelve languages, FLORES+
FLORES+ results for the 12 Tier 1 languages, a second test set.
Builder questions, answered.
Not here? Ask on GitHub or read the docs.
Which models use the tokenizer today?
Gemma 3 models, after resizing the embedding to 294,145 rows. A separate pack extends Llama 3.1.
Does a model trained with Nzinga improve?
Not measured yet. A controlled experiment is built and tested: the same Gemma 3 model trained with and without Nzinga. Results will be published either way.
How are members paid for their recordings?
Per accepted minute of their own voice, at a rate set per market at launch. When partners license recordings, a share of what they pay comes back in proportion to accepted minutes.
Where does my audio go today?
It stays on your phone. Only the details are sent: language, length and your choice. When upload is built, the plan is encrypted storage with keys per member.
Is Nzinga hosted in Africa yet?
Not yet. Code, releases and compute run on GitHub, Hugging Face, Modal and Cloudflare. The roadmap moves training data and production serving to Africa, layer by layer.
How do I get an invite to join?
Ask a member: everyone starts with three invites and earns one more per accepted hour. Each code works once. Or try the demo with code 482917.
Count your own sentence in your browser.
Paste a sentence in your language and see the tokens it needs, against twelve other tokenizers.