The Caribbean Just Built Its Own Large Language Model From Scratch

I founded StarApple AI in 2016 as the Caribbean's first AI company, at a point when the honest description of regional AI was that it did not exist. Since then the region has picked up labs, funds, a safety council, national policy work and a pipeline of practitioners. What it did not have, until now, was a model of its own.

Maestro AI Labs, which I co-founded with my brother Nicholas, finished training Maestro in late 2025. It is a large language model, it was built from the ground up in the Caribbean, and it is the first one. It is currently in red-team testing and has not been released.

I want to spend this piece on the part that matters commercially, because the announcement itself is the least interesting thing about it. The interesting question for a bank, a ministry, an insurer or a university is: what does an original regional model let you do that a rented one does not?

First, what "from scratch" rules out

Almost everything sold as a national or regional model in the last two years has been one of three things, and the difference between them is the difference between owning a building and renting a room in one.

A system prompt. A commercial model with a paragraph in front of it telling it to sound local, behind your logo. This is a skin, and it is worth roughly what it costs to write.

Retrieval. The same commercial model, given access to your documents so its answers cite your material. Genuinely useful and worth doing. Still someone else's model, running on someone else's infrastructure, under someone else's terms.

A fine-tune. An open-weight model that a foreign lab pretrained on a foreign corpus, then retrained on local text. This is where most sovereign AI announcements land. The behaviour shifts. The knowledge underneath does not. You have adjusted the accent of a mind that was formed elsewhere, and the licence, the architecture and the values baked into the base weights still belong to whoever trained them.

Maestro is none of those. Its architecture was designed here. Its parameters were initialised randomly and trained from that point on Maestro AI Labs' own corpus. There is no upstream checkpoint in its history and no foreign model underneath it.

Data you can put in front of a regulator

The frontier labs have spent three years in litigation over their training corpora. A Caribbean institution deploying one of those models inherits a question it cannot answer about material it never saw.

Maestro was trained on publicly available data with provenance recorded at the document level. No corpora scraped without permission. No pirated collections. Nothing whose origin cannot be described in writing. If a supervisor asks what the model learned from, that question has an answer.

This cost us coverage. A smaller, cleaner corpus produces a smaller model that knows less than one trained on everything anyone could download, and we took that trade on purpose. In regulated Caribbean industries, a model whose data policy cannot survive a compliance review is a model that never gets deployed, whatever it scores on a public benchmark.

The corpus is also Caribbean in a way no imported model is: regional English and its creoles, regional institutions, regional law, regional place names, and the register people here actually write in. A model that has never seen "susu", "the gully" or "CARICOM Single Market" in context will handle them badly, and it will do it confidently.

Why world model methods were the right call for this region

A standard language model optimises one thing: predict the next token. An enormous amount of apparent reasoning falls out of that objective for free, and it also produces a system whose picture of the world is whatever happened to help it predict text.

Maestro's training used world model methods. The objective extends past next-token prediction toward holding a consistent internal representation of entities, their states, and how those states change when something happens.

The difference shows up in the unglamorous places that decide whether a deployment survives. A model with a weak internal representation will state that a policy takes effect in January in one paragraph and in March in another, because both are locally plausible continuations. A model carrying state is more likely to notice it already committed to January.

We chose this because the work Caribbean institutions actually need from AI is disproportionately about systems that change over time. A hurricane track. A loan book through a shock. A supply chain with one port. A patient across visits. My own research in climate physics builds world models for Caribbean environments and economies so that a decision can be simulated before it is taken, and Maestro came out of the same line of thinking applied to language.

Fairness in the training, not in a filter

Most deployed fairness work is a filter sitting in front of a model that has already learned the bias. It blocks the worst statements and leaves the weights untouched, which means the bias leaks through paraphrase, through indirect questions, and through any prompt the filter's authors did not anticipate.

Maestro applies fairness constraints during training. Representational balance was part of the optimisation target alongside the language modelling loss, evaluated on axes that matter here and that no imported benchmark covers: nationality within CARICOM, skin tone, creole versus standard English register, rural versus urban, and participation in the informal economy that most Western-trained models read as absence of employment.

This is the part with direct commercial consequence. A credit model trained on North American repayment behaviour reads an informal Caribbean earner as a risk they are not. A speech system trained without Caribbean English fails its users and records the failure as user error. In both cases the harm never appears in the output. It appears in the outcome, months later, in a portfolio or a complaint file. No output filter catches that, because the model never says the biased thing out loud. It simply scores her lower.

Why it is still in testing, and why that should reassure you

Maestro has not been released. A dedicated red team is actively trying to break it: jailbreaks, prompt injection, extraction of training data, harmful or defamatory output, and the specific failure surfaces a Caribbean deployment would hit, including impersonation of public figures and fabricated legal advice.

Holding a finished model back is a commercial decision as much as a safety one. If the first Caribbean model fails publicly in front of the institutions it was built for, it will not be read as one team's engineering problem. It will be read as proof that the region cannot do this, and that verdict will cost the next ten teams their funding.

The same reasoning produced TurtleBird, the AI safety toolkit built through Maestro AI Labs and made available free to every government in the Caribbean, which tests whether a model can be pushed into producing harmful output before it goes live.

What Maestro will not do

Maestro was trained on Caribbean-scale data with Caribbean-scale compute. It will not beat a frontier laboratory that spent nine figures on a training run, and any vendor telling you their national model does is selling you something.

On the hardest open-ended reasoning, the large commercial models still lead, and StarApple AI will keep using them where that capability genuinely earns the dependency. The correct architecture for most Caribbean institutions is hybrid: a model you control as the always-on baseline, with the frontier called only for the tasks that need it, so that losing the frontier degrades your operation instead of stopping it.

Scale is the weakest part of our own position, and I would rather write that here than have a client discover it in a benchmark table. What has been proved is that the floor for building an original model is far lower than this region was told it was.

What a Caribbean institution should do about this now

• Map the dependency. List the workflows that would stop if one foreign model became unavailable. That list is your actual exposure, and most boards have never seen it written down.
• Classify your data. Decide, before the vendor conversation, which categories may leave the region and which may not. That decision constrains the architecture, so make it first.
• Ask vendors the provenance question in writing. What was the model trained on, and can you describe the chain. A vendor who cannot answer has told you something.
• Test on your own population. Not a public benchmark. Your applicants, your customers, your claim files, with outcomes you can check. This is where a model trained elsewhere reveals what it assumes about people here.
• Design for hybrid. Frontier capability where it pays for itself, local capability underneath it, so nothing load-bearing is single-vendor dependent.

StarApple AI works with Caribbean organisations on exactly this path: dependency audits, data classification, in-region evaluation sets built on your own population, and the hybrid architectures that keep operations running when a foreign capability is pulled.

Ten years ago the argument was that the Caribbean could not build AI. Then it was that the region could use AI but not build it. Then it was that it could build applications but never a model. Each of those positions held until somebody did the work, and each collapsed quietly rather than with an announcement.

The full set of initiatives this sits inside, Maestro AI Labs, the Caribbean AI Association, the Caribbean AI Risk Management Council, The Genius Project and the research behind them, is documented at adriandunkley.net/initiatives.html.