NatureV1: global weather and hurricane track/intensity forecasting on the iHELIX engine, verified on held-out seasons.
AI & ML interests
Based in Jamaica. We build AI models from the ground up.
Recent Activity
Organization Card
CorX Labs is Jamaica's independent AI research lab. We build Jamaican AI from the ground up: five open models so far, from CorX3.8, a 27-billion-parameter Jamaican Patois assistant, to NatureV1, a weather and hurricane model. Most were trained from random weights on a single GPU, and where one is a fine-tune, we say so.
| Models built | Released with weights | Technical reports | Licence |
|---|---|---|---|
| 5 | 4 | 5 | Apache-2.0 |
Models
| Model | What it is | Size | Built from | Report |
|---|---|---|---|---|
| CorX3.8-27B | Jamaican Patois conversational assistant. Patois by default, English on request. | 27B | Fine-tune of Qwen3.8-27B, fully merged | 03 |
| CorX1.5 | Decoder-only language model with learned role embeddings | 157.8M | From scratch, tokenizer included | 01 |
| TriStream-SVS | Singing-voice synthesis: pitch, timbre and breath in three separate streams | 321.8M | From scratch | 02 |
| NatureV1 | Probabilistic weather and hurricane model, attention in kilometres on the sphere | 89.45M | From scratch, on the iHELIX engine | 05 |
| HELIX | Linear-scaling sequence architecture that keeps exact long-range recall | configurable | From scratch; no trained checkpoints | 04 |
Browse by collection: Text · Singing voice · Weather
Papers
Read them on the web or download the PDFs: CorX Labs Papers
 |
 |
 |
 |
 |
- CorX1.5: a 158M-parameter language model trained from random weights on one GPU. (PDF) The architecture, the three-stage training, and the memory arithmetic that fits three billion tokens through one card.
- TriStream-SVS: building the source-filter model of the voice into a singing synthesiser. (PDF) Three streams that cannot share information until a late fusion trunk, with zero-shot cloning and a 32-step flow-matching decoder.
- CorX3.8-27B: giving an open 27B model a Jamaican voice. (PDF) What a rank-64 LoRA on 19 million tokens changed, and what it did not.
- HELIX: separating memory capacity from bandwidth in a linear-scaling sequence model. (PDF) N1.00 attention pairs, and recall that ties full attention while the same model without its index falls to chance.
- Hurricane tracks without a neural network. (PDF) A physics consensus that beats extrapolation and CLIPER at every lead on 4,447 forecasts from four seasons it never saw.
Measured, not promised
| Hurricane track error (km), 2018 and 2020-2022, all basins | 24 h | 48 h | 72 h | 120 h |
|---|---|---|---|---|
| Straight-line extrapolation (strongest per lead) | 179 | 442 | 751 | 1,417 |
| CLIPER | 163 | 378 | 589 | 914 |
| NatureV1 hurricane physics, no neural network | 127 | 301 | 515 | 884 |
Global AI models such as GraphCast and Pangu-Weather remain far better beyond 12 hours, and the report says so. The same rule runs through every model here: results are measured on data the model never trained on, "trained from scratch", "fine-tuned" and "not trained yet" are kept apart, and nothing is claimed that has not been measured. TriStream-SVS, for example, has no listening tests yet, so it reports no quality scores.
corx-labs.com · Chat with CorX3.8 · Documentation · Blog · Benchmarks · PyPI: naturev1, ihelix, helix-lm
CorX Labs · Westmoreland, Jamaica · Architecting the future, from the ground up.
datasets 0
None public yet