Who worked on it.
CorX Labs is a one-person research lab. All three models — CorX3.8-27B, TriStream-SVS and CorX1.5 — plus the training pipelines, the documentation and this site, were built by the developer below.
Every CorX Labs model was built by Nathan, a 14-year-old developer from Jamaica and the founder of the lab. He designed the architectures, trained the tokenizer, wrote the pretraining and fine-tuning pipelines, curated the data and ran every training job — solo, on a single GPU. That covers CorX3.8-27B, Jamaica's first large open-weight LLM; TriStream-SVS, a singing voice synthesis model; and CorX1.5, trained from random weights with no base model underneath it. There is no team.
Nathan
Founder & AI Researcher · Jamaica · Age 14
Nathan started building with AI models before he could build one. At twelve he was experimenting with APIs, wiring up a chatbot on top of Gemini and customising its behaviour himself. It worked — but it was somebody else's model wearing his configuration, and that gap bothered him.
At thirteen he moved to fine-tuning, and it went badly. Slow GPUs, training runs that took longer than his patience, and eventually the decision to stop. That failure is the reason CorX1.5 exists: the next attempt skipped adaptation entirely and started from random weights.
CorX1.5 is the result — a custom decoder-only transformer with role embeddings and assistant-only loss masking, a 32,768-token tokenizer trained from scratch, and a three-stage pipeline covering pretraining, distillation and reasoning. Open-sourced under Apache 2.0, so anyone can download it and take it apart the same way he had to.
Worked on
- Model architecture — RoPE, SwiGLU, RMSNorm, role embeddings, assistant-only loss masking
- Custom 32,768-token BPE tokenizer
- Data curation across 25 open datasets, and the ~3B-token pretraining pipeline
- Distillation and automated quality filtering
- Reasoning training, including the step-by-step
<think>format - The hardcoded crisis-response safety layer
- Model card, documentation and ongoing checkpoint releases
Shipped
Three models, one author.
Everything CorX Labs has released, and what Nathan actually did on each one. The distinction between "trained from scratch" and "fine-tuned" is stated plainly rather than blurred.
| Model | What it is | Size | Built from | Released |
|---|---|---|---|---|
| CorX3.8-27B | Jamaican Patois conversational assistant — Jamaica's first large open-weight LLM | 27B | Fine-tune of Qwen3.8-27B, fully merged | 2026 |
| TriStream-SVS | Singing voice synthesis — pitch, timbre and breath in three parallel encoder streams | 321.8M | From scratch | 2026 |
| CorX1.5 | Decoder-only transformer language model, trained from random weights on one GPU | 157.8M | From scratch, including the tokenizer | 2026 |
CorX3.8-27B is a fine-tune, and CorX1.5 is not. CorX3.8-27B starts from the open-weight Qwen3.8-27B base, so its reasoning, code and maths ability come from that base; what Nathan added is the Jamaican Patois voice, the register and the cultural grounding. CorX1.5 started from random weights — architecture, tokenizer, pretraining and fine-tuning all written here. Both are worth having; they are not the same achievement, and the lab does not present them as one.
The journey
Three attempts, one model.
The short version of how a teenager in Jamaica ended up with a language model trained from random weights.
-
Age 12
APIs and a chatbot
Built a working chatbot on top of Gemini's API and customised it himself — prompts, behaviour, personality. Genuinely useful, and genuinely somebody else's model. The question that stuck: what would it take to make the model itself?
-
Age 13
Fine-tuning, and giving up
Moved to fine-tuning existing base models. The hardware could not keep up — training runs dragged on for far too long on slow GPUs, and eventually he stopped. This is the part of the story most people leave out of their timeline. It is also the part that decided the next one.
-
Age 14
From the ground up
Instead of adapting a model, he wrote one. A custom decoder-only transformer, a tokenizer trained from scratch, ~3 billion tokens of pretraining across 25 open datasets, then distillation and reasoning training. Debugged crash by crash on a single GPU. The result is CorX1.5.
-
Now
Three models, all open
CorX1.5 was followed by TriStream-SVS, a singing voice synthesis model built around the source-filter theory of the human voice, and then by CorX3.8-27B — a 27B Jamaican Patois assistant and Jamaica's first large open-weight LLM. All three are published on Hugging Face and none of them is finished: new training data and targeted fixes go in as gaps turn up in testing, with updated checkpoints released as they are ready.
“For years I've shared a dream to build something great — my own AI. Today that dream came through, and I can say I made it from scratch, from the ground up. So I'm open-sourcing it for people like you to download.”
— Nathan, on releasing CorX1.5
Questions
About the developer
Who built CorX3.8-27B?
Nathan, the founder of CorX Labs — a 14-year-old developer and independent AI researcher from Jamaica. CorX3.8-27B is a fine-tune of the open-weight Qwen3.8-27B base, fully merged and standalone; the Jamaican Patois voice, register and cultural grounding are his work, while the underlying reasoning, code and maths ability come from the base model.
Who built CorX1.5?
Nathan built CorX1.5 alone, from random weights. He designed the decoder-only transformer architecture, trained a 32,768-token BPE tokenizer from scratch, curated roughly 3 billion tokens of pretraining data across 25 open datasets, and ran the distillation and reasoning stages — all on a single GPU. It is not a fine-tune of any existing base model.
How many people work at CorX Labs?
One. CorX Labs is a one-person research lab in Jamaica. Every model, the training pipelines, the documentation and this website were built by Nathan.
How old is the founder of CorX Labs?
Nathan is 14. He started with AI model APIs at twelve, attempted fine-tuning at thirteen and abandoned it when the hardware could not keep up, and wrote CorX1.5 from scratch at fourteen.
Which models has Nathan released?
Three: CorX3.8-27B, a 27B Jamaican Patois assistant and Jamaica's first large open-weight LLM; TriStream-SVS, a 321.8M parameter singing voice synthesis model that factors pitch, timbre and breath into three parallel encoder streams; and CorX1.5, a 157.8M parameter transformer trained from random weights. All are published openly on Hugging Face.
Can I contribute to a CorX model?
Yes. The most useful contributions right now are finding where a model breaks and reporting it, pointing at open datasets worth training on, or proposing an evaluation that would show something real at these model sizes. Email lkk89002@gmail.com.
Contributing
Want to work on a CorX model?
Use it and report back
The most useful contribution right now is finding where the model breaks. Run it, break it, and send what you found.
Data and evaluations
Open datasets worth training on, or an evaluation that would show something real about a model this size — both are welcome.
Just get in touch
Collaboration, mentorship, compute, or a question — one inbox, read by the person who wrote the code.