Instructions to use K-Ai-Official/K-Oss-9-Mini with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use K-Ai-Official/K-Oss-9-Mini with Transformers:

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

pipe = pipeline("text-generation", model="K-Ai-Official/K-Oss-9-Mini")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("K-Ai-Official/K-Oss-9-Mini")
model = AutoModelForCausalLM.from_pretrained("K-Ai-Official/K-Oss-9-Mini", device_map="auto")
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=256)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use K-Ai-Official/K-Oss-9-Mini with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "K-Ai-Official/K-Oss-9-Mini"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "K-Ai-Official/K-Oss-9-Mini",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/K-Ai-Official/K-Oss-9-Mini

SGLang

How to use K-Ai-Official/K-Oss-9-Mini 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 "K-Ai-Official/K-Oss-9-Mini" \
    --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": "K-Ai-Official/K-Oss-9-Mini",
		"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 "K-Ai-Official/K-Oss-9-Mini" \
        --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": "K-Ai-Official/K-Oss-9-Mini",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Unsloth Desktop
Docker Model Runner
How to use K-Ai-Official/K-Oss-9-Mini with Docker Model Runner:
```
docker model run hf.co/K-Ai-Official/K-Oss-9-Mini
```
Browse Quantizations to use this model in llama.cpp, Ollama, LM Studio, or any compatible app.

K-Oss-9-Mini

Next-Generation Multilingual AI by K-Ai

K-Oss-9-Mini is a next-generation language model developed by K-Ai (Kortex Ai), succeeding the previous generation, K-Oss-8-Mini-Instruct-8B.

K-Oss-9-Mini represents the next generation of the K-Oss model family.

Model Generation

Generation	Model
Previous Generation	K-Oss-8-Mini-Instruct-8B
Next Generation	K-Oss-9-Mini

The new generation introduces a dedicated dataset, Kai_V2, alongside a structured training, evaluation, and reproducibility pipeline.

Model Details

Property	Value
Developer	K-Ai (Kortex Ai)
Model Name	K-Oss-9-Mini
Generation	Next Generation
Architecture	Qwen3
Base Model	Qwen3-8B-Base
Training Method	QLoRA Fine-Tuning
Framework	Unsloth, 【entity-Transformers¦canonical_name=Transformers】, TRL
Quantization During Training	4-bit
Target Languages	Indonesian, English, Japanese
Training Dataset	Kai_V2
Model Type	Causal Language Model
Intended Tasks	Text Generation

Base Model

K-Oss-9-Mini uses:

Qwen/Qwen3-8B-Base

The original pretrained architecture and knowledge come from the Qwen team (Alibaba Cloud).

K-Ai builds upon this foundation through its own fine-tuning methodology, dataset development, evaluation process, and model integration.

Training Methodology

The model development pipeline uses:

Component	Description
QLoRA	Parameter-efficient fine-tuning with a 4-bit quantized base model
Unsloth	Optimized model loading and training
TRL SFTTrainer	Supervised fine-tuning
PEFT / LoRA	Trainable adapter layers
Response-only training	Training loss is applied to assistant responses rather than user prompts

Training configuration, optimizer settings, random seed, package versions, and hardware information should be recorded in a reproducibility manifest.

Dataset

K-Oss-9-Mini is developed using Kai_V2, a multilingual dataset maintained by K-Ai.

Dataset repository:

K-Ai-Official/Kai_V2

Kai_V2 covers eight major categories:

Category	Focus
Instruction Following	Understanding and following instructions
Reasoning	Logical and analytical tasks
Coding	Programming and software development
Science	Scientific knowledge
Language Learning	Multilingual education
Japanese Kaiwa	Practical Japanese conversations
Creativity	Creative generation
Safety	Responsible AI behavior

The dataset provides separate training, validation, and benchmark splits.

The benchmark split is kept separate from model training to reduce evaluation leakage.

Supported Languages

Language	Code
Indonesian	`id`
English	`en`
Japanese	`ja`

Multilingual Benchmark

K-Oss-9-Mini is designed for multilingual evaluation across Indonesian, Japanese, and English.

The benchmark evaluates the model's ability to understand instructions, generate accurate responses, perform reasoning, and communicate naturally in different languages.

Supported Benchmark Languages

Language	Code	Evaluation Focus
Indonesian	`id`	Natural communication, reasoning, instruction following
Japanese	`ja`	Japanese comprehension, grammar, vocabulary, and Kaiwa
English	`en`	Instruction following, reasoning, coding, and general knowledge

Benchmark Categories

Category	Description
Instruction Following	Ability to understand and follow instructions
Reasoning	Logical thinking and problem-solving
Coding	Programming knowledge and code generation
Science	Scientific understanding and explanations
Language Learning	Grammar, vocabulary, and language comprehension
Japanese Kaiwa	Practical Japanese conversational ability
Creativity	Creative writing and generation
Safety	Responsible responses and refusal behavior

Evaluation Method

Each language is evaluated using language-appropriate prompts and expected outputs.

Evaluation methods include:

No.	Method
1	Exact normalized matching
2	Keyword and phrase matching
3	Required keyword coverage
4	Code execution tests
5	Expected refusal checks
6	Manual and rubric-based evaluation

Benchmark Integrity

Principle	Description
Separate benchmark	Benchmark dataset is maintained separately from training and validation data
No training on benchmark	Benchmark samples must not be used for training or fine-tuning
Reduced leakage	Separation helps reduce data leakage and supports more reliable evaluation

Evaluation Transparency

Topic	Note
Reporting	Benchmark results should be reported separately for each language and category
Japanese content	Includes language-learning examples and practical everyday conversations
Language coverage	Depends on available training data and should not be interpreted as equal proficiency across all three languages

Intended Use

K-Oss-9-Mini is intended for:

Use Case
Multilingual AI experimentation
Instruction-following applications
Educational assistance and language learning
Coding assistance and programming experiments
Research into compact language models
AI application prototyping

Limitations

Limitation
This model inherits the limitations of its Qwen3 base model.
Fine-tuning does not guarantee improved performance in every task.
Output quality depends on dataset size, diversity, and review.
Japanese language accuracy requires further evaluation.
The model may generate incorrect or misleading information.
Safety behavior requires additional testing.
It is not intended for high-stakes medical, legal, or financial decisions.

Evaluation

Model evaluation uses the independent Kai_V2 benchmark.

The evaluation pipeline includes:

No.	Step
1	Greedy decoding
2	Automated rule-based checks
3	Instruction-following evaluation
4	Multilingual output inspection
5	Basic refusal behavior checks
6	Optional comparison against the original base model

No performance scores or superiority claims are reported without actual evaluation results.

Inference

Example using 【entity-Transformers¦canonical_name=Transformers】:

from 【entity-transformers¦canonical_name=Transformers】 import AutoTokenizer, AutoModelForCausalLM
import torch

model_id = "K-Ai-Official/K-Oss-9-Mini"

tokenizer = AutoTokenizer.from_pretrained(model_id)

model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.float16,
    device_map="auto"
)

messages = [
    {
        "role": "user",
        "content": "Introduce yourself in Indonesian."
    }
]

inputs = tokenizer.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    return_tensors="pt"
).to(model.device)

outputs = model.generate(
    inputs,
    max_new_tokens=256,
    do_sample=False
)

response = tokenizer.decode(
    outputs[0][inputs.shape[-1]:],
    skip_special_tokens=True
)

print(response)

This example assumes the merged model and its tokenizer have been successfully published and verified.

Reproducibility

K-Ai's development pipeline records:

Item	Description
Base model revision	Version / revision of Qwen3-8B-Base used
Dataset checksums	Checksums for Kai_V2 splits
Training configuration	Hyperparameters and training config
Random seed	Seed used for training
GPU information	Hardware used for training
Python and package versions	Environment versions
Training logs	Logs from training runs
Evaluation outputs	Benchmark and evaluation results

This information supports reproducibility and transparent model development.

License

Apache License 2.0.

K-Oss-9-Mini is based on Qwen3-8B-Base, which is released under Apache-2.0.

Please retain the original model's attribution and applicable license notices when redistributing.

K-Ai (Kortex Ai)

We Build Because We Believe We Can.

Built on the Qwen3 architecture, K-Oss-9-Mini is developed through QLoRA fine-tuning to advance K-Ai's work in compact multilingual AI models.

The model focuses on Indonesian, English, and Japanese, with capabilities targeting instruction following, reasoning, coding, scientific knowledge, language learning, and everyday conversations.

«We Build Because We Believe We Can.»

Downloads last month: 443

Safetensors

Model size

8B params

Tensor type

BF16

Model tree for K-Ai-Official/K-Oss-9-Mini

Base model

Qwen/Qwen3-8B-Base

Quantized

unsloth/Qwen3-8B-Base-unsloth-bnb-4bit

Finetuned

(172)

this model

Quantizations

1 model

K-Ai-Official
/

K-Oss-9-Mini