Announcement
CorX Labs has built five open-source AI models in Jamaica, including the island's first large language model and a hurricane model.
Jamaica · First released · Updated
CorX Labs, an independent artificial intelligence research lab and research station based in Jamaica, has built five open-source AI models. Four are released with trained weights: CorX3.8-27B, a 27-billion-parameter Jamaican Patois-speaking assistant and the island's first large open-weight LLM; CorX1.5, a 158-million-parameter language model designed and trained entirely from scratch; TriStream-SVS, a 321.8-million-parameter singing voice synthesis model; and NatureV1, an 89-million-parameter weather and hurricane model.
The fifth, HELIX, is a linear-scaling sequence architecture published as code. The lab also makes Automa, an Android app for running a company of AI agents from a phone. All of it is the work of a single developer.
CorX3.8-27B — a model that speaks Patois
CorX3.8-27B is a 27-billion-parameter conversational assistant that speaks Jamaican Patois by default and switches to standard English on request. It answers directly, holds its boundaries, and carries everyday Jamaican cultural reference rather than treating Patois as a novelty accent applied to a foreign voice.
The lab states plainly that this one is not built from scratch. CorX3.8-27B is a fully merged fine-tune of the open-weight Qwen3.8-27B base (Apache 2.0): the underlying reasoning, code and mathematics come from that base, and CorX Labs contributed the Jamaican voice, register and grounding. It is also not a Jamaican knowledge base — it should not be trusted for specific facts such as addresses, current laws or prices without a retrieval layer, which the lab ships to ground answers on real Jamaican sources including OpenStreetMap and public government documents. It runs in 4-bit on a single ~24GB GPU, or about 54GB in full bf16.
CorX1.5 — built from random weights
CorX1.5 is not a fine-tune of an existing base model. Its architecture, its 32,768-token tokenizer, its pretraining pipeline and its fine-tuning were all written in house and trained from random weights. The model is a custom decoder-only transformer using RoPE positional encoding, SwiGLU feed-forward layers and RMSNorm, with 10 layers and a 1,024-token context window, and it introduces learned role embeddings — a signal that marks every token as user, assistant or raw text at the architecture level, rather than leaving the model to infer it from special tokens.
Training ran in three stages: pretraining on approximately 3 billion tokens drawn from
25 open datasets spanning code, mathematics, reasoning, conversation and general text;
distillation from a larger teacher model with automated quality filtering; and reasoning
training, in which the model learns to show its working inside
<think> blocks before answering. A held-out validation split with
early stopping was used throughout.
The lab is publishing the model with an explicit statement of its limits. CorX1.5 is a small model trained on billions of tokens where comparable models from major labs are trained on trillions, and CorX Labs states plainly that it is not suitable for broad factual recall, complex multi-step mathematics, or production and safety-critical use. The model also ships with a hardcoded crisis-response layer, implemented in code rather than left to the model's judgment.
TriStream-SVS — a voice factored into three streams
TriStream-SVS is a 321.8-million-parameter singing voice synthesis system that factors the source-filter model of the human voice into three parallel encoder streams: a source stream carrying pitch alone, a filter stream carrying timbre and formants — the only stream that carries singer identity — and a residual stream for breath and fricative texture. Separating identity into a single stream is a deliberate design choice: it makes the part of the model that determines who the voice sounds like an isolated, inspectable component.
TriStream-SVS is released and available on Hugging Face. The lab has not published objective benchmark scores or listening-test results for it, and is not claiming any it has not measured.
NatureV1 — a hurricane model, trained and scored
NatureV1 is an 89-million-parameter AI model of the atmosphere. It rolls the global atmosphere forward three days and, for a tropical cyclone, forecasts its track as weighted scenarios, its intensity, the chance of rapid intensification and the chance of landfall — every number with an uncertainty. Its attention heads work in kilometres over the real sphere rather than in cells on a flat grid, and because hurricane labels are scarce, the backbone is frozen before the storm heads train, so only about 1.2 million parameters ever see one. It trained in about a week on a single GPU.
Every result was measured on data the model never trained on. On 59 storms from the held-out 2017, 2019 and 2021 Atlantic seasons, its hurricane track beats a physics-only steering track and straight-line extrapolation at every lead — 163 km at 24 hours and 739 km at 72 — and on the unseen 2023 El Niño season it beat steering for every one of the seven hurricanes.
The lab is just as explicit about the gap: those errors are 2.5 to 4.5 times the National Hurricane Center’s, the model underestimates strong storms, and it missed Hurricane Otis’s explosive intensification entirely. It is a research model, not an operational forecast, and must not be used for real decisions.
HELIX — linear scaling, exact recall
HELIX is a sequence architecture that drops attention's quadratic cost without giving up
exact long-range recall, by braiding a local window, a recurrent state and a landmark-tree
index over one residual stream. It is published as the helix-lm package with a
tested reference implementation; there are no trained checkpoints. NatureV1's engine grew
out of it.
“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.”
Nathan, founder of CorX Labs
Availability. Every model is open-source under the Apache 2.0 licence:
- CorX3.8-27B — huggingface.co/Sigmandndnns/CorX3.8-27B (merged weights). You can talk to it in the browser at corx-labs.com/chat.
- CorX1.5 — huggingface.co/Sigmandndnns/CorX1.5-160M, including weights, tokenizer and the model definition required to load it.
- TriStream-SVS — huggingface.co/Sigmandndnns/TriStream-SVS-300M (released).
- HELIX — pypi.org/project/helix-lm (reference implementation; no trained checkpoints).
- NatureV1 — huggingface.co/Sigmandndnns/NatureV1-500 (weights, results and evaluation scripts) and pypi.org/project/naturev1.
- Automa — Android app; no public download yet. How it works.
Training is ongoing across the line-up, with updated checkpoints released as new data is added.
About CorX Labs. CorX Labs is an independent AI research lab and research station in Jamaica, in the Caribbean. It designs, trains and open-sources AI models built from the ground up. Media and business inquiries: lkk89002@gmail.com.