Fewer tokens for African languages.
Gemma 3's tokenizer plus 32,000 African word pieces. Every existing token keeps its ID, and English barely changes.
Four things the premium costs you.
The premium is paid on every query for a model's life. Cutting it at the tokenizer costs little.
Lower cost
At any per-token price, a message that needs fewer tokens costs less to run.
Longer context
A model reads a fixed number of tokens, so more text fits.
Speed
Models write one token at a time, so replies arrive sooner.
Lifetime
A tokenizer is fixed into every model trained with it.
Load it in one line, keep every ID.
Load the JSON file with the Hugging Face tokenizers library. To use it with Gemma 3, resize the embeddings to 294,145 rows and train the new ones.
Measured, not promised.
FLORES+ devtest, 1,012 sentences per language, 3 November 2026. Method in the docs.
“Mawaƙin Rap yana canza sunan - zuwa Ye.”
Tokenizer, explained.
Still unsure? Paste your own sentence into the live comparison.
Which models fit it?
Gemma 3 models today. A Llama 3.1 pack exists, and Gemma 4 needs a rebuild.
Does it make a model better yet?
Not measured yet. A controlled experiment, the same Gemma 3 model trained with and without Nzinga, is built and waits for GPU time.
How was it tested?
On FLORES+ and NTREX-128, which were never used in training. Raw token counts for the same sentences, against twelve other tokenizers.
What happens to English and French text?
Almost nothing. Both stay within one to three tokens of Gemma 3, on tens of thousands, in both test sets.
Try it in your language.
Paste a sentence and see the tokens it needs.