We stopped calling ourselves a publishing company.
Because it stopped being true, and the thing that replaced it is more interesting.
For most of the last year, the fastest way to explain AfriStack was by pointing at AfriKDP. It worked. People understood a publishing platform, they understood 85% royalties, and they understood why African authors needed payment rails that actually reach them.
It is no longer an accurate description of the company, and continuing to use it would be a slow form of dishonesty. So this issue is the correction.
AfriStack Smart System Ltd is an AI and technology infrastructure company. AfriKDP, Skilluxo and AfriAssist are the products where that technology is deployed, measured and improved.
That is not a rebrand. It is a description of where the engineering time has actually gone.
What changed
Somewhere between shipping the second product and the third, the interesting problem stopped being the products. Three platforms all needed the same things: a model that could understand a document, generate a draft, translate a passage, classify a request, answer a question about a body of text. Building that capability three times would have been absurd. Building it once, properly, underneath all of them, was the obvious move — and once you commit to that, you are an infrastructure company whether or not you have updated your website.
We have now updated our website.
Omnis: we trained our own model
The part of this that is hard to say without sounding like everyone else is that we are building our own AI. So let us skip the adjectives and give you the numbers.
Omnis Tiny 0.1 is a 42-million-parameter decoder-only language model, trained from scratch. No pretrained weights. No external inference API in the loop. The tokenizer, the architecture, the training loop, the data pipeline and the evaluation harness are ours.
It ran for 15,000 steps over 983 million tokens, in roughly twelve hours, on a single 6 GB consumer laptop GPU.
| Metric | Step 500 | Final |
|---|---|---|
| Held-out perplexity | 237.5 | 41.2 |
| Validation loss | 5.4700 | 3.7179 |
| Next-token top-1 accuracy | 19.9% | 34.5% |
| Next-token top-5 accuracy | 36.9% | 55.9% |
| Factual continuation, top-10 | 0% | 75% |
| Median fact rank | 209 | 7 |
Then we instruction-tuned it: 504 steps over 63,000 conversations, about fifteen minutes of training, final validation loss 1.6479. The model stopped continuing text and started answering — and, importantly, it holds its own identity rather than claiming to be somebody else's assistant, which is what happens by default when you train on web text.
What it cannot do
This is the part most companies leave out, so we will put it in the middle rather than the footnotes.
A 42-million-parameter model is not a product. It cannot reliably follow complex instructions, reason, or answer factual questions. It falls into repetition loops on some prompts. Its train/validation gap tells us we hit the useful ceiling for this parameter count on this corpus. Its world knowledge is thin — it knows that Shakespeare wrote Romeo and Juliet with high confidence, and it does not reliably know the capital of France.
Answer format is correct. Answer content is often wrong. That is exactly what a model this size should do, and anyone who tells you otherwise is selling something.
Omnis Tiny is a learning vehicle. Its entire job was to prove that our pipeline works end to end — data, tokenizer, architecture, training, evaluation, inference — and it did that. Every failure we hit at this size cost us hours. The same failure at a serious scale would cost six figures.
What is actually running in production
Our live AI features do not run on Omnis Tiny. They run on Omnis Base — our adaptation track, where we take a strong open-weight model with a commercially usable licence and train low-rank adapters against benchmarks derived from our own products rather than generic leaderboards. Five adapter generations so far, each one a measured step.
Self-hosting rather than renting an API is a deliberate choice. It means we control the cost, the latency, the availability, and where customer data goes — the last of which is not a nice-to-have for the universities and institutions already publishing through AfriKDP.
The architectural decision that makes all of this work
If you remember one thing from this issue, make it this one.
No AfriStack product ever calls a model directly. Every product calls the Omnis Gateway, and the gateway decides which model answers.
Inside AfriKDP, every AI feature is individually routable. One environment variable moves a single feature to a different backend; another moves all of them at once. If a backend is unavailable, features fall back automatically instead of failing. And our admin tooling reports requests, failures, truncation and fallbacks per provider, per feature.
That measurement is the whole point. It turns "we should use our own model here" from a conviction into a threshold. When Omnis clears the bar on a capability, we move that capability — and serving it gets dramatically cheaper. Until then, we do not move it, no matter how much we would like the story.
Where we intend to be unmatched
We are not going to match frontier labs on general English reasoning. That door is closed, and it is closed to almost everyone — the compute alone is a nine-figure problem before you count the research team.
The door that is open is narrower and considerably more defensible. Frontier models are weak on Yorùbá, Igbo, Hausa and Pidgin. They are weak on local curricula, local business norms and local context. Not because those problems are hard, but because the data is not on the open web and the market is too small for them to prioritise. That weakness is structural, it is measurable, and it will persist.
A one-billion-parameter model that is the best in the world at Yorùbá is a better business than a thirty-billion-parameter model that is fortieth-best at English. That is the target we are aiming at.
What this means for you
- If you use our products — nothing breaks. AfriKDP, Skilluxo and AfriAssist keep shipping, and they get more intelligent as the layer underneath them improves.
- If you are an investor — the thing you are looking at is no longer three apps. It is an AI company that already has three distribution channels and a working measurement loop between them.
- If you are a partner or developer — the AfriKDP Publishing API is live today. AI services and platform APIs follow as Omnis matures.
Next issue: what we actually shipped across all three products — twenty-two AI capabilities in AfriKDP, a full book-writing studio, and an agent system in AfriAssist.