Sovereignty, measured layer by layer:where Nzinga stands.

Nzinga's purpose is sovereignty, not only fewer tokens: African control of the tokenizer, the data, the standard it is measured against and the infrastructure it runs on. Here is where each layer stands today.

The tokenizer

10.9% of the vocabulary, 32,000 of 294,145 tokens, is designed from Nzinga's own corpus; the rest is Gemma 3's. On Tier 1 text, 21.7% of the tokens used come from Nzinga's own vocabulary.

The data

Training text exists for all 12 Tier 1 languages, from open web text and Google WAXAL transcripts. None of it yet comes with community consent or from African partners. From the scorecard:

own_vocabulary: 32,000 of 294,145
own_token_share_tier1: 21.7%
tier1_with_training_text: 12 of 12
tier1_within_target: 6 of 12
model_families: gemma3, llama31

Layer three: the performance standard

The Tier 1 premium averages 1.55 times English; 6 of 12 languages meet the 1.3 to 1.5 target. The accuracy gap is not yet measured.

Layer four: infrastructure and people

Nothing is hosted in Africa yet: GitHub, Hugging Face, Modal and Cloudflare. The people layer is open too. Today:

• No native-speaker reviewer yet, in any language
• No benefit-sharing agreement with communities yet
• No published change policy for releases
• Tokenizer bound by Gemma terms

“Indicators are kept separate, because any weighting would be a judgement call.”