CorX3.8-27B is a 27-billion-parameter conversational language model from CorX Labs in Jamaica. It is a Jamaican Patois-speaking assistant that answers directly, holds boundaries and carries everyday Jamaican cultural reference, while keeping the general reasoning, coding and mathematics ability of its base model. It is fine-tuned from the open-weight Qwen3.8-27B base under Apache 2.0, fully merged and standalone.

Is CorX3.8 trained from scratch?

No. Unlike CorX1.5, which was trained from random weights, CorX3.8 is a fine-tune of an existing open-weight base model, Qwen3.8-27B, released under Apache 2.0. The underlying reasoning comes from that base; CorX Labs added the Jamaican voice, the direct register and the cultural grounding.

Does CorX3.8 know specific Jamaican facts like addresses or laws?

Not reliably. CorX3.8 has a Patois voice and common cultural reference, but it does not dependably know specific facts such as street addresses, current laws, exam schedules, prices or business regulations. For those, pair it with the recommended retrieval layer, which grounds answers on real Jamaican sources and says plainly when something is not in its references — do not trust its unaided recall.

What hardware does CorX3.8-27B need?

It runs in 4-bit quantisation on a single GPU with roughly 24GB of memory. Full bf16 precision needs about 56GB. The weights are fully merged, so there is no adapter step and no separate base-model download.

Released Jamaican Patois 27B params Apache 2.0 Open weights

CorX3.8-27B

Q: What licence is CorX3.8 released under?

Apache 2.0, inherited from the Qwen3.8-27B base model, with attribution to Qwen retained as the licence requires. The weights are open, so anyone can run, study and build on it.

Jamaica's first large open-weight LLM — a Jamaican Patois-speaking assistant that answers straight, holds its ground, and carries a natural Jamaican voice. Third in the CorX line, after CorX1.5 and TriStream.

Chat with it View on Hugging Face

Repository: Sigmandndnns/CorX3.8-27B · Developer: Nathan · Origin: Westmoreland, Jamaica

CorX3.8-27B is a 27-billion-parameter conversational model from CorX Labs in Westmoreland, Jamaica. It speaks Jamaican Patois by default and switches to standard English when you write in English or ask it to. It is a fine-tune of the open-weight Qwen3.8-27B base (Apache 2.0) — fully merged and standalone, with no adapter step and no separate base download — so it keeps its base model's general reasoning, code and maths while carrying a Jamaican voice, a direct register and everyday cultural grounding.

Parameters: 27B
Base model: Qwen3.8-27B
Languages: Patois & English
Format: Fully merged
Runs on: ~24GB GPU (4-bit)
Full precision: ~56GB (bf16)
Licence: Apache 2.0
Status: Released

The character

How it behaves.

The fine-tune adds a voice and a temperament on top of the base model's reasoning — this is what CorX Labs shaped.

Patois by default

It speaks Jamaican Patois out of the box, and switches to standard English the moment you write in English or ask it to. The voice is the point, not a costume you have to request.

Answers direct

No filler, no throat-clearing, no false agreement. It gives its honest read of a question and owns its mistakes rather than papering over them.

Holds boundaries

It declines requests that cross a line and says so plainly, instead of complying under pressure. Refusal is stated, not negotiated away.

Culturally grounded

Everyday knowledge a Jamaican would take for granted, expressed in a natural voice — the reference points and idiom that make a reply feel local rather than translated.

Honest about the build

A fine-tune, and we say so.

Being upfront about limits is part of the CorX character, so the model page follows the same rule. CorX3.8 is not trained from scratch. It is a fine-tune of Qwen3.8-27B, an existing open-weight base released under Apache 2.0. The underlying reasoning, code and mathematics come from that base — CorX Labs added the Jamaican voice, the direct register and the cultural grounding on top. That makes it a different kind of project from CorX1.5, which was trained from random weights.

Capabilities at a glance

CorX3.8-27B capability levels
Area	Level
Jamaican Patois conversation	Strong
Direct, honest register	Strong
Common Jamaican cultural reference	Good
General reasoning & explanation	Good (inherited)
Code & mathematics	Good (inherited)
Specific facts (roads, laws, prices)	Weak — use retrieval
Real-time / current information	None — no live data

What it is not

Not trained from scratch. The reasoning is inherited from Qwen3.8-27B; the Jamaican voice and register are what CorX Labs added.
Not a Jamaican knowledge base. It has the voice and common reference, but does not reliably know addresses, current laws, exam schedules or prices — pair it with retrieval.
Not fluent in every register. Its Patois is strong for conversation and common idiom; deep dialect range across all fourteen parishes is a goal for a future release.
Not a substitute for professional advice — legal, medical or financial. It can explain and discuss; it is not an authority.

Grounding

For facts, ground it on real sources.

Because the model should not be trusted for specific facts on its own, CorX Labs ships a retrieval layer that grounds answers on real Jamaican sources — OpenStreetMap geography and public government documents. With retrieval attached, CorX3.8 answers factual questions from cited material and says plainly when something is not in its references, instead of guessing. This is the recommended setup for anything factual.

Usage

Running CorX3.8-27B.

The weights are fully merged, so it loads with the standard transformers API — no adapter step, no separate base download.

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

tok = AutoTokenizer.from_pretrained("Sigmandndnns/CorX3.8-27B")
model = AutoModelForCausalLM.from_pretrained(
    "Sigmandndnns/CorX3.8-27B",
    torch_dtype=torch.bfloat16,
    device_map="auto")

msgs = [{"role": "user", "content": "Who are you?"}]
prompt = tok.apply_chat_template(msgs, tokenize=False, add_generation_prompt=True)
inputs = tok(prompt, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=256, temperature=0.7)
print(tok.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

Runs in 4-bit on a single ~24GB GPU; full bf16 needs ~56GB. Apache 2.0, inherited from the base model, with attribution to Qwen retained as the licence requires.

Open weights

Download it, run it, build on it.

The merged weights are published under Apache 2.0 on Hugging Face — open for anyone to run, study, and build on.

View on Hugging Face Back to all models