CorX Labs

An AI lab in the Caribbean.

CorX Labs is an independent artificial intelligence research lab and research station based in Jamaica. We build language models from the ground up.


CorX Labs designs, trains and open-sources its own language models. Not fine-tunes of somebody else's base model — models written from zero: the architecture, the tokenizer, the data pipeline, the training loop. Everything the lab releases is published under the Apache 2.0 licence so anyone can download it, read it, and learn from it.

Founded
2026
Based in
Jamaica
Models released
1
Licence
Apache 2.0

The story

Built by one person, from zero.

CorX Labs was founded by Nathan, a solo developer from Jamaica. For years the goal was the same one: build an AI — not adapt someone else's, but write the thing himself and train it up from random weights.

It started with APIs. An early chatbot built on top of Gemini, customised piece by piece — which worked, but was never the same as building from the ground up. Then came fine-tuning, which failed: slow GPUs, training runs that took forever, and eventually the decision to walk away from that approach.

CorX1.5 is what came out the other side. A custom transformer architecture, a tokenizer trained from scratch, a three-stage pipeline covering pretraining, distillation and reasoning — designed, debugged and trained solo on a single GPU. Every checkpoint in the repository represents real hours spent fixing data pipelines, chasing down overfitting, and learning how these systems actually work underneath.

It is open-sourced for exactly that reason: so other people can take it apart the same way.

Founder

Nathan

Founder & AI Researcher · Jamaica

Wrote the architecture, trained the tokenizer, ran the pretraining and fine-tuning, and maintains the model repository. Reachable at lkk89002@gmail.com.

What we stand for

Small, transparent, and fully understood.

From the ground up

If we ship it, we wrote it. Understanding every part of a system is the point of building it, not a side effect.

No overclaiming

A 158M-parameter model is a 158M-parameter model. We document the limits as carefully as the capabilities.

Open by default

Apache 2.0, published on Hugging Face, documented here. Research that nobody can inspect isn't research.

Work with us

Business inquiries welcome.

Collaboration, questions about a model, or something you want to build with us — the door is open.