CorX Labs

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 @CorXLabs]

Announcing CorX Labs: An
Independent AI research station based in jamaica.

We design, train, and opensource our own language models completely from scratch from the tokenizer to the training loop.

Read about our first model release, CorX1.5:
https://t.co/lpEaQUVYQx

[CorX Labs @CorXLabs]

Added ai benchmarks and leaderboards.

https://t.co/hxZJnGTS3K

[CorX Labs @CorXLabs]

School will be reopening back Monday,
(I'm still in high school)
We won't get to post much, but we are actively making models -- (nathan)

#jamaica #school

[CorX Labs @CorXLabs]

CorX3.8 now features full endpoint integration. The model can autonomously execute external API calls, conduct real-time web searches, and process live data to augment its responses.

​As demonstrated in image, this enables dynamic context retrieval while maintaining our native localized reasoning capabilities.
​Available now: [https://t.co/NFDekCPEmY

Please note we have set up a temporary live demo for testing purposes. You can access it here: https://t.co/ydga7TSRxz
]

[CorX Labs @CorXLabs]

An advanced agentic AI platform engineered to autonomously reason, plan, and execute multi-step workflows. It dynamically interacts with external APIs and environments to accomplish complex objectives from end-to-end with minimal human intervention.

​As this is an early-stage live demo, expect bugs while development continues.

[CorX Labs @CorXLabs]

#chat #jamaica #ai

[CorX Labs @CorXLabs]

https://t.co/JutbIUi6Ag

[CorX Labs @CorXLabs]

https://t.co/JutbIUi6Ag

[CorX Labs @CorXLabs]

https://t.co/JutbIUi6Ag

[CorX Labs @CorXLabs]

CorX3.8-27B.

#jamaica #ai #llm

[CorX Labs @CorXLabs]

Jamaica’s first large open-weight LLM is live: CorX3.8-27B.

​Built in Westmoreland by CorX Labs, CorX3.8-27B brings natural Jamaican Patois fluency, local cultural grounding, and a direct, honest register to a 27B parameter model.

​What it brings to the table:
• Native Voice: Speaks Patois by default and shifts seamlessly to English when prompted.

• Direct Register: Zero filler, no corporate throat-clearing or false agreement—just clear, honest answers.

• Cultural & Technical Depth: Grounded in Jamaican context while retaining strong reasoning, math, and coding ability.

• Standalone & Open: Fully merged weights under Apache 2.0. No adapters or extra base steps required.
​Run it in 4-bit on a single 24GB GPU or bf16 on ~56GB.

​(Fine-tuned on Qwen3.8-27B, adding Jamaican Patois alignment, local context, and direct register to a strong open-weight base.)

​🔗 Try it on Hugging Face:

https://t.co/4mS1p5f1u7

​#Ai #OpenSource #JamaicanPatois #CorXLabs #CaribbeanTech #LLM

[CorX Labs @CorXLabs]

TriStream-SVS a 300M Singing voice model We built completely from scratch, source-filter factored architecture and all got featured by @HuggingModels.

still training, open source under Apache2.0.

This is surreal 🇯🇲

#jamaica #opensource #ai

[Hugging Models @HuggingModels]

Ever wished for an AI that can handle multiple data streams at once? Meet TriStream-SVS-300M, a compact yet powerful model that processes three streams simultaneously. It's Apache 2.0 licensed, so you can use it freely. 564 downloads already, and it's just getting started! #AI #MachineLearning

[Hugging Models @HuggingModels]

With 300M parameters, this model balances performance and efficiency. It's designed for finetuning, meaning you can adapt it to your specific needs without massive compute. The 'TriStream' architecture hints at triple-input handling, making it versatile for complex data fusion. #Tech #DeepLearning

[Hugging Models @HuggingModels]

TriStream-SVS-300M is a base model for finetuning, perfect for custom tasks. Think of it as a Swiss Army knife for multimodal or multi-source data. You can build apps that analyze text, audio, or sensor data together. Ideal for real-time analytics, smart assistants, or research prototypes. #AI #BuildInPublic

[Hugging Models @HuggingModels]

Why choose TriStream-SVS-300M? It's Apache 2.0, so no licensing headaches. The base_model tag means it's ready for transfer learning. With 564 downloads and growing, the community is taking notice. Whether you're a researcher or developer, this model offers a solid foundation for innovation. #AI #OpenSource

[Hugging Models @HuggingModels]

Ever wished for an AI that can handle multiple data streams at once? Meet TriStream-SVS-300M, a compact yet powerful model that processes three streams simultaneously. It's Apache 2.0 licensed, so you can use it freely. 564 downloads already, and it's just getting started! #AI #MachineLearning

[CorX Labs @CorXLabs]

#svs #jamaica #ai #opensource #tristream #321M #carribean #fromscratch

[CorX Labs @CorXLabs]

Today we are introducing

TriStream-SVS, the Caribbean's first open-weight Singing Voice Synthesis model developed completely from scratch.

TriStream-SVS (321M parameters) introduces a structural voice separation architecture that explicitly decouples human voice into three distinct pipelines: Pitch, Timbre/Identity, and Breath.

This structural wiring ensures the model retains exact speaker characteristics without bleeding into musical notes.

Zero-shot voice cloning via a highly optimized Speaker

Encoder.Natural human expression modeling, incorporating micro-pitch corrections, scoops, and a custom swing parameter to prevent mechanical grid alignment.Rectified

Flow Matching architecture, achieving high-fidelity audio generation in just 32 steps

a massive leap in efficiency over traditional diffusion methods.

Please note that this model is currently in active development. TriStream-SVS is a 321M parameter generative audio model that utilizes Rectified Flow Matching to synthesize expressive human singing. By structurally decoupling audio into three distinct streams

an early-stage model trained on limited regional datasets and hardware constraints, users should manage expectations regarding its current output variance.

Full technical documentation:
https://t.co/foLLmu1iWK

[CorX Labs @CorXLabs]

https://t.co/H5aLPct7Pi