Instructions to use mkd-hossain/keural-14.8b-sft with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use mkd-hossain/keural-14.8b-sft with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="mkd-hossain/keural-14.8b-sft")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("mkd-hossain/keural-14.8b-sft")
model = AutoModelForCausalLM.from_pretrained("mkd-hossain/keural-14.8b-sft")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps

vLLM

How to use mkd-hossain/keural-14.8b-sft with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "mkd-hossain/keural-14.8b-sft"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "mkd-hossain/keural-14.8b-sft",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/mkd-hossain/keural-14.8b-sft

SGLang

How to use mkd-hossain/keural-14.8b-sft with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "mkd-hossain/keural-14.8b-sft" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "mkd-hossain/keural-14.8b-sft",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "mkd-hossain/keural-14.8b-sft" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "mkd-hossain/keural-14.8b-sft",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use mkd-hossain/keural-14.8b-sft with Docker Model Runner:
```
docker model run hf.co/mkd-hossain/keural-14.8b-sft
```

Keural 14.8B — SFT (Instruction Tuned)

Keural is a 14.83B parameter bilingual Korean-English Mixture-of-Experts language model trained from scratch. This repository contains an early SFT checkpoint (7,000 steps) fine-tuned on the mkd-chanwoo/keural-SFT dataset in HuggingFace Mixtral-compatible safetensors format.

Note: This is an early SFT checkpoint (7,000 / 18,000 steps). Full SFT training is still in progress. Final SFT and DPO-aligned chat model will be released as mkd-hossain/keural-14.8b-chat.

Model Architecture

Property	Value
Architecture	Mixtral MoE (MixtralForCausalLM)
Parameters	~~14.83B total (~~2.9B active per token)
Layers	24
Hidden size	4096
Attention heads	32 (GQA: 8 KV heads)
Experts	8 total, top-2 active per token
FFN intermediate size	5632
Context length	4096 tokens
Vocabulary	131,072 + 2 special tokens (131,074)
RoPE theta	500,000
dtype	bfloat16

Training History

Stage	Steps	Tokens	Details
Pretraining	100,000	~64.56B	Korean + English web text
Annealing	20,000	~5.16B	High-quality filtered data
SFT (this checkpoint)	7,000	~450M	ChatML instruction tuning

SFT Dataset

Trained on mkd-chanwoo/keural-SFT:

1,134,119 samples across 14 curated sources
~710M tokens total
Korean 44% / English 56%
Sources: UltraChat, OpenOrca, KoAlpaca, MathInstruct, AIHub, GSM8K, Magicoder, and more
Format: ChatML (<|im_start|> / <|im_end|>)

Tokenizer

Token	String	ID
BOS	`<bos>`	1
EOS	`<eos>`	2
PAD	`<pad>`	0
UNK	`<unk>`	3
IM_START	`<\|im_start\|>`	131072
IM_END	`<\|im_end\|>`	131073

Usage

vLLM (Recommended)

pip install vllm==0.9.2 --no-build-isolation
pip install "transformers==4.57.0"
vllm serve mkd-hossain/keural-14.8b-sft --dtype bfloat16 --max-model-len 4096

Transformers

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model_id = "mkd-hossain/keural-14.8b-sft"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

prompt = """<|im_start|>user
What is artificial intelligence?
<|im_end|>
<|im_start|>assistant
"""

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(
    **inputs,
    max_new_tokens=256,
    temperature=0.7,
    top_p=0.9,
    repetition_penalty=1.1,
    do_sample=True,
)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Roadmap

Stage 1 Pretraining — 100K steps, ~64.56B tokens
Stage 2 Annealing — 20K steps, ~5.16B clean tokens
SFT — 7K/18K steps (in progress)
SFT — Full 18K steps
DPO alignment
Keural Chat model release (mkd-hossain/keural-14.8b-chat)

Citation

@misc{keural2026,
  title  = {Keural: A Bilingual Korean-English MoE Language Model},
  author = {MKD Hossain},
  year   = {2026},
  url    = {https://huggingface.co/mkd-hossain/keural-14.8b-sft}
}

Trained from scratch on KT Cloud NIPA2-H200 infrastructure using FSDP distributed training on 2× NVIDIA H200 GPUs.

Downloads last month: 108

Safetensors

Model size

15B params

Tensor type

BF16

Model tree for mkd-hossain/keural-14.8b-sft

Unable to build the model tree, the base model loops to the model itself. Learn more.