CorX Labs
Documentation
Documentation

The things we have made.

The models CorX Labs releases are open-source and published on Hugging Face, and the apps we build sit alongside them. Each one has a page here with what it is, what it can do, and where it stops.

Apps

Our apps

Android Agent orchestration Device Files No public download yet

Automa

Run your AI company from your phone. The agents run on a server — this phone under Termux, or a machine over HTTPS — and Automa is the control surface: what needs you, what they are working on, and the files they are waiting on.

Platform
Android
Shape
Client for an agent server
Default port
3100
File channel
Device Files
Recurring work
Routines
Download
Not linked yet
See how Automa works Ask about it

Released

Our models

Released Weather & hurricanes 89.45M params Probabilistic Research model

NatureV1

Jamaica’s open-source AI weather and hurricane model. Its attention works in kilometres on the real sphere, not in cells on a flat grid; it rolls the atmosphere forward three days and, for a hurricane, forecasts track, intensity, rapid intensification and landfall, every number with an uncertainty attached.

Parameters
89.45M
Blocks
17 iHELIX
Head radii
120–1,600 km
Input grid
1.5°, 89 channels
Weather rollout
To 72 h, skilful
24 h track error
163 km, held-out
View model details View on Hugging Face
Released Jamaican Patois 27B params Apache 2.0

CorX3.8-27B

Jamaica's first large open-weight LLM — a Jamaican Patois-speaking assistant that answers straight, holds its ground, and carries a natural Jamaican voice. A fine-tune of Qwen3.8-27B, fully merged and standalone.

Parameters
27B
Base model
Qwen3.8-27B
Languages
Patois & English
Runs on
~24GB GPU (4-bit)
Format
Fully merged
Licence
Apache 2.0
View model details View on Hugging Face
Released Singing voice synthesis 321.8M params Zero-shot cloning

TriStream-SVS

A singing voice synthesis model that builds the source-filter theory of human voice production into its architecture as a hard constraint — three parallel encoder streams for pitch, timbre and breath, combined only in a late fusion trunk.

Parameters
321.8M
Encoder streams
3 parallel
Generation
Rectified flow matching
Sampling steps
~32
Sample rate
24 kHz
Status
Released
View model details View on Hugging Face
Architecture proposal Linear scaling Long context No trained checkpoints

HELIX

Hierarchical Episodic Linear IndeX — a sequence architecture that drops attention’s quadratic bill without giving up exact long-range recall. Three strands per block: an exact local window, a recurrent state, and an index that can reach any block in the past through a landmark tree.

Package
helix-lm 0.1.0
Strands per block
3
Attention pairs
N1.00 measured
Hot decode state
O(1)
Probes per step
O(log N)
Status
Untrained reference implementation
View architecture details View on PyPI
Released Text generation 158M params Apache 2.0

CorX1.5

A ~158M parameter decoder-only transformer, built entirely from scratch — architecture, tokenizer, pretraining and fine-tuning, all written and trained solo on a single GPU. Not a fine-tune of an existing base model.

Parameters
157.8M
Layers
10
Context
1,024 tokens
Vocabulary
32,768 BPE
Pretraining
~3B tokens
Licence
Apache 2.0
View model details Try it on Hugging Face
In progress

More is in progress.

CorX Labs is a working research lab, not a finished catalogue. New checkpoints and new models are published as they are ready — the release feed and the Hugging Face commit history are the fastest way to see what has changed.

Documentation

How these models are built.

The architecture choices, the three-stage training pipeline, the safety layer and the loading instructions — all in one place.

Read the documentation Contact us