Pixel mosaic in Nzinga colours, one bright cell at its core
Docs
/
Funders

For Funders and Investors

The problem, what is built and measured, the plan to prove the model, and the decisions ahead.

Measured
Plan
01

The Team

Nzinga is built by HQ Protocol and GTCX, two teams building sovereign AI and African-led infrastructure.

Lwazi Wali · Founder

Venture builder and impact investor. Previously Head of Venture at Founders Factory Africa, founder of Her Quarters, a social capital market for African women founders, and founder of the Impact Investment Exchange South Africa chapter. Earlier, lead investment officer at a major South African investment holding company, after seven years in aerospace and energy finance. 2019 Obama Foundation Leader, Africa.

Amani Anai · Technical Architect

Economist (UC Berkeley) and Lead Architect of GTCX Protocol, building market infrastructure for African commodities. Previously Design Partner at Dfi, working on regenerative economics.

02

The Token Premium

AI models read, bill and limit text in tokens. For the same meaning, African languages need more tokens than English. That premium is paid on every query, in cost, capacity and speed.

Average token premium against English
FLORES+
OpenAI cl100k (GPT-4)
2.71x
Llama 3.1
2.67x
OpenAI o200k (GPT-4o)
2.09x
Gemma 3 and Gemma 4
1.97x
Nzinga
1.55x
12 Tier 1 languages · 1,012 sentences each · tokenizer README

A tokenizer is fixed into every model trained with it. Its premium is paid for the life of that model.

03

What Is Built and Measured

evidence
today
Fewer tokens than 12 other tokenizers on all 12 Tier 1 languages
measured
22% fewer tokens than the best of them
measured
6 of 12 Tier 1 languages within the 1.3x to 1.5x target
measured
Fewer tokens on 24 of 26 languages on NTREX-128
measured
Consent rules, invites and earnings ledger in shared code
built
Phone app with a demo mode
built
Model accuracy with Nzinga
not yet measured
Sources: tokenizer README, Nzinga README
Not built

Audio upload, payouts through a payment partner, a database and calls. Yoruba and Akan remain above the token target.

04

The Proof-of-Concept Plan

The plan is twelve weeks. Collect consented, paid voice recordings in one language from about 100 invited members. Turn them into a versioned dataset with a consent record per item. Measurably improve a speech model. License it to one design-partner buyer, with members paid through one real payout rail. Read the full plan: web version or PDF.

success criteria
planned
Collection works: accepted hours
set with partner
Members stay: recording in week 4
at least half
Consent holds: valid record per speaker
100%
Quality holds: share accepted after review
tracked
Data is useful: word error rate vs base model
a measured drop
Money flows: paid within the window
100%
Someone pays: signed pilot licence
1
Source: proof-of-concept plan, section 0
Note

The word error rate is reported whatever it is.

05

Sovereignty Milestones

Sovereignty is measured layer by layer, each with a baseline and a target. Where each layer stands today:

sovereignty scorecard
today
Control of the tokenizer: 10.9% own vocabulary
target 100%
Control of the data: 0% community-consented
target majority
Equal performance: 1.55x; 6 of 12 on target
target all 12
Infrastructure: none hosted in Africa
target in Africa
People and benefit: 0 reviewers, no agreement
target all 12
Source: SOVEREIGNTY.md, Nzinga tokenizer repository

The voice network adds its own indicators, proposed until adopted. Consent records per speaker and withdrawal are built. African hosting, a governance body and the licence income split are not yet in place.

The full scorecard is on the Trust page.

06

The Business Model

Nzinga is a three-sided system. Each side needs a different promise.

three sides
planned
Members: recordings, reviews, invites
pay + share
Data buyers: licence fees, collection budgets
datasets
Organisations: seats, projects, licences
own-language tools
Source: proof-of-concept plan, section 1

Nzinga's own models use recordings under "Nzinga's models". Better models lift the value of everything else.

Decision

Where licence income goes is to be decided. The plan splits it four ways: a member pool, a community fund, operations and Nzinga. No shares or prices are set.

07

Decisions and Milestones Ahead

decisions for the owner
in order
The pilot language and the design-partner buyer
1
Pilot market, payout provider, where raw audio may be stored
2
Member rate and the split of licence income
3
Licence terms and banned uses
4
Whether Nzinga for Teams is in the pilot
5
Governance body for the library and the fund
6
When calls enter scope
7
Source: proof-of-concept plan, section 18
tokenizer milestones
planning estimates
Model experiment: accuracy with and without Nzinga
next
Corpus and governance, data partnerships
months 1 to 3
Rebuild on Gemma 4; own vocabulary decision
months 3 to 6
v1.0, independently reproduced
months 6 to 9
Source: tokenizer ROADMAP.md · estimates, not commitments

To discuss funding or investment, contact the team.