The Caribbean AI built from Scratch
Built in Jamaica
Jamaica Built Its Own Language Model. That Changes Something.
In Kingston, a small Caribbean team has done work usually associated with the world’s wealthiest technology companies and research laboratories: training a language model from the ground up.
By Caribbean AI Newsletter · Kingston, Jamaica · 3 September 2026 · Updated 18 September 2026
In 2026, a team in Kingston set machines to work on something unusual for the region: learning statistical relationships across language through the team’s own training process.
Maestro AI Labs, a Jamaican AI research group, trained a language model from the ground up with contributions from researchers and professionals across the Caribbean. It is the lab’s second version: the first adapted existing models, and the team used what it learned there to build this one from scratch. The model is now undergoing security testing. It has no public name, release date or announced access terms.
A Caribbean team completed the process of constructing and training an underlying language model instead of only building an interface on top of one made abroad. For a region that has spent most of the generative AI era consuming systems developed somewhere else, that is a meaningful threshold.
The immediate question is whether the project can be documented, shared and improved well enough to become useful research infrastructure for the region.
Building the thing changes your relationship with it
Most people experience artificial intelligence from the top. You type into a box, words return and the machinery underneath remains invisible. Building a model reverses that relationship. Before there is a chatbot, a team has to make decisions about data, tokenization, architecture, training objectives, optimization, compute, checkpoints, evaluation and failure.
The application is close to the end of the process. The model itself is the product of thousands of prior technical decisions, each of which can shape what it knows, how it fails and which groups it represents poorly.
A simplified route from regional material to a testable model
Much of the Caribbean AI market operates at the application layer. Companies build useful software around models created by OpenAI, Anthropic, Google, Meta, Mistral and others. That is legitimate engineering and a separate job from producing the model weights. Fine-tuning is another valid path, in which an existing foundation model receives additional training for a narrower domain or behaviour.
Training a base model means taking responsibility for more of the machine: the data mixture, architecture, optimization process, learned behaviour, weaknesses and eventual mistakes. Maestro has now worked in two of these categories. Its first version, reported by the Jamaica Observer in April 2026, adapted existing models and retrained them in a Jamaican context; the second version, covered here, belongs in the third category because it was trained from the ground up.
Three different things that are often described as “our AI”
Application wrapper
An interface sends a user’s request to another company’s model. The underlying intelligence comes from an upstream provider, whatever value the product adds on top.
Application layerFine-tuned model
A team starts with an existing foundation model and continues training it with specialized examples. The original model remains underneath.
Adaptation layerBase model training
A team runs the training process that produces the model weights. This demands more data engineering, compute, experimentation and machine-learning expertise.
Model layerThe Caribbean now has a machine that learned here
Language models are shaped by the material they encounter during training. Ask a frontier model about New York, London or Silicon Valley and it operates in information-rich territory. Ask increasingly specific questions about Caribbean institutions, communities, linguistic patterns, laws, businesses, places or cultural references and the information environment can become thinner.
These systems sometimes perform well and sometimes improvise. Many Caribbean users have experienced the moment when a sophisticated model becomes strangely unfamiliar with the place they live.
A regional training programme can move Caribbean information closer to the beginning of the model-building process. Names, institutions, geography, language and cultural context become part of the material from which the system constructs its representation of the world, rather than additions supplied only after a foreign model has already been trained.
A possible instrument for bias research
Caribbean organizations increasingly depend on language models built outside the region. Researchers can test whether those systems respond differently to regional names, dialects, locations, occupations, financial situations or institutional contexts. They can measure a disparity in outputs, but explaining its precise cause is harder when the full training mixture is unavailable.
A documented regional model could provide another reference point. Researchers could present the same Caribbean-focused tests to several systems and compare the results. Because the local system’s own data and method could be inspected, it may help researchers form and test explanations for differences observed in closed systems. That would give the region a controlled research instrument for studying proprietary models from the outside.
The Allen Institute for AI demonstrates what deep openness can look like through its Olmo model programme. Ai2 publishes access to training data, model weights, code, evaluation tools and parts of the model-development flow. A small Jamaican laboratory can apply the same principle at its own scale, because the research value of any model grows with what other teams can inspect.
Why a small model can still be strategically useful
The AI industry often presents progress as a contest in scale: more parameters, GPUs, tokens, context and compute. A compact regional model will lose many direct capability comparisons with frontier systems. Complex reasoning, large software projects, advanced agents, very long contexts and difficult scientific work may remain better suited to the strongest available commercial or open frontier models.
But smaller models can offer a different set of benefits. Depending on model size and optimization, an organization may run one on local infrastructure, control where information travels, reduce variable inference costs and continue operating in environments with limited connectivity. A government department classifying routine records does not necessarily need the most capable model in the world for every request. A bank or hospital may value a narrow system that stays inside its own technical boundary.
The practical architecture is likely to be hybrid
Regional small model
- Repeatable classification and tagging
- Answers grounded in a fixed document set
- Local-language and fairness experiments
- Private or lower-connectivity deployments
Vetted frontier model
- Hard multi-step reasoning
- Large coding and agentic workflows
- Very long or multimodal inputs
- High-stakes work with formal review
This makes “sovereign AI” less abstract. Local training builds technical capability even when the chips, software frameworks and much of the infrastructure come from global suppliers: engineers learn to manage training runs, researchers see how data choices affect behaviour, institutions learn what infrastructure they need and students gain a system they may eventually be able to inspect.
Maestro’s own history shows how that capability compounds. When the Jamaica Observer covered the first version in April 2026, a core team of three people, backed by AI agents, volunteers and young Jamaican developers, was building it while a dedicated red team probed the system for weaknesses. The team carried the lessons from that first version into a second model built entirely from scratch. Caribbean AI expert Adrian Dunkley, who led the project, describes the build as having two purposes.
“We built Maestro to show the youth that we can build it here, and to show the world what an LLM built completely by us can do.”Adrian Dunkley, Maestro AI Labs
The experiment began with physics
One of the project’s more unusual choices concerns its modelling philosophy. Dunkley, whose academic background includes mathematics and physics, says parts of the training approach were influenced by how physicists represent systems through constraints, state variables and relationships.
Physics models do not permit arbitrary behaviour. Conservation laws, boundary conditions and symmetries restrict the set of plausible solutions. Machine learning has borrowed from this idea in fields such as physics-informed neural networks, where known relationships are included in the learning objective or model structure.
Maestro explored related ideas while developing its language-model method. That could become the project’s most scientifically interesting contribution, because physics-informed methods are well established in scientific machine learning but less common in language modelling.
Jamaica’s experiment has a larger Latin American counterpart
On 10 February 2026, Chile’s National Center for Artificial Intelligence, CENIA, formally launched Latam-GPT, an open regional language-model initiative for Latin America and the Caribbean. CENIA says the collaboration involved more than 60 institutions across 15 countries and more than 100 professionals.1
CENIA reports that Latam-GPT was developed on a Llama 3.1 70-billion-parameter base architecture, complemented by a regional corpus and evaluation benchmarks. It also reports more than 300 billion tokens collected under explicit permissions and licences.1 The Maestro model is a much smaller project: a second version trained from scratch through the lab’s own model-building process.
How Latam-GPT and the Maestro model compare
| Question | Latam-GPT | Maestro AI Labs model |
|---|---|---|
| Status | Publicly launched on 10 February 2026. | Second version completed in 2026 and now in security testing; no release date announced. |
| Leadership | Coordinated by CENIA with public, academic and technology partners. | Privately led by Maestro AI Labs in Jamaica. |
| Scale | 60+ institutions, 15 countries and 100+ professionals, according to CENIA. | Caribbean researchers and professionals contributed; count not published. |
| Technical base | Llama 3.1 70B base architecture plus a regional corpus and regional benchmarks, according to CENIA. | Trained from the ground up, drawing on lessons from a first version that adapted existing models; architecture and parameter count not published. |
| Data disclosure | CENIA reports 300+ billion tokens obtained under explicit permissions and licences. | Trained on publicly available sources; source list, proportions and volume not published. |
| Primary emphasis | Open regional capability across Latin America and the Caribbean, primarily Spanish and Portuguese contexts. | Anglophone and Creole-speaking Caribbean research, local capability and bias evaluation. |
Latam-GPT shows what regional coordination can accomplish when governments, universities, development institutions and technology partners work at scale. The Jamaican project asks what a much smaller Caribbean team can learn by controlling more of its own model-building process. The region gains from having both: continental infrastructure in Latam-GPT and a focused laboratory experiment in Kingston.
CENIA director Álvaro Soto described Latam-GPT as a base technology from which the region can create its own solutions.1 The same logic applies at a smaller scale in Jamaica, where the knowledge and people produced by Maestro’s first version made the second one possible.
What happens next will determine the project’s value
The next stage is about making the work usable by others. A technical paper should identify the architecture, parameter count, training objective, data composition, compute environment, evaluation design, bias tests, security findings and known failure modes. A clear data statement should name source categories and explain the legal basis for their use, since public availability and permission to train are separate legal questions.
Bias research is one of the project’s stated purposes, so the evaluation design deserves as much care as the model. A shared test set should cover regional names, dialects, geography, employment, financial services, education, public services and cultural assumptions, with results reported for every model tested, including where each one fails.
The strongest outcome would be a public Caribbean benchmark that regulators, universities and companies can run against any model. That benchmark could outlast this particular system and give the region a shared method for asking whether imported AI works fairly in Caribbean contexts.
Five markers that the experiment is becoming infrastructure
- A technical paperA complete method description would let researchers reproduce the physics-inspired approach and extend it.
- Research accessAccess terms that let Caribbean universities and research organizations test capability, regional knowledge, bias and safety for their own use cases.
- A documented data statementSource categories, proportions, licences, exclusions and governance decisions should be published.
- A public regional benchmarkReusable Caribbean-focused evaluations may become more valuable than any single model.
- A third versionMaestro moved from adapting existing models to training its own in 2026. A stronger third version would show that Maestro can keep improving at that pace.
Frequently asked questions
Sources and evidence notes
- CENIA, “Latam-GPT: la primera IA regional abierta creada con datos latinoamericanos,” 10 February 2026. Source for the launch date, base architecture, collaboration scale, dataset size and licensing description.
- Allen Institute for AI, Olmo. Source for Ai2’s description of its open model flow, training data and development tools.
- Ai2 documentation. Source for Ai2’s statement that it releases training code, data, weights and recipes.
- Kellaray Miles, “Jamaican AI loading,” Jamaica Observer, 15 April 2026. Source for the April 2026 description of Maestro’s first version, its core team and red-team testing.
- Details specific to the second Maestro AI Labs model, including its from-scratch build and the quotation from Adrian Dunkley, were supplied by its development team.