K-Ai-Official/K-Oss-8-Mini-Instruct-Beta

K-Oss 8 Mini Instruct Beta

K-Oss 8 Mini Instruct Beta is an experimental, compact member of the K-Oss 8 model family developed by K-Ai (Kortex AI).

It is a multimodal (text + image) instruction-tuned model designed for English, Indonesian, and Japanese (日本語).

The model was fine-tuned from "Qwen/Qwen3-VL-8B-Instruct" using QLoRA. This repository hosts the resulting LoRA adapter weights and requires the original base model to run.

«Release Status: Experimental Beta

This release is intended for experimentation, pipeline validation, and early evaluation.

The adapter was trained for 5 steps on 36 examples. Its primary purpose is to verify the fine-tuning pipeline and introduce K-Ai model identity behavior.

Expect behavior close to the original base model. This release has not undergone comprehensive training or benchmarking and is not recommended for production deployment.»

The K-Oss 8 Family

• K-Oss 8: 120B total parameters, 12B active (MoE, A12B), based on NVIDIA Nemotron 3.5 Super VL.
• K-Oss 8 Mini Instruct Beta: A compact experimental member of the family, built on Qwen3-VL-8B-Instruct.

K-Oss 8 Mini Instruct Beta is a separate, smaller model. It does not use Nemotron weights and is not distilled from K-Oss 8.

It shares the K-Oss family identity, development goals, and multilingual focus, but it should not be assumed to match the capabilities of the 120B model.

Model Details

Property| Description Developer| K-Ai (Kortex AI) Model Name| K-Oss 8 Mini Instruct Beta Model Type| LoRA adapter for a vision-language model Base Model| Qwen/Qwen3-VL-8B-Instruct Base Parameters| 8,810,770,672 Trainable Parameters| 43,646,976 (0.50%) Input| Text and images Output| Text Target Languages| English, Indonesian, Japanese Release Channel| Experimental Beta License| Apache-2.0

Training Procedure

Method: QLoRA (4-bit base weights via bitsandbytes, through Unsloth) with supervised fine-tuning using TRL "SFTTrainer" and the Unsloth vision data collator.

LoRA Configuration

• Rank ("r"): 16
• Alpha: 16
• Dropout: 0.0
• Bias: None
• Trained modules: Attention and MLP modules of the language model.
• Frozen modules: Vision tower, preserving the base model's visual capabilities while reducing memory requirements.
• Gradient checkpointing: Unsloth.

Hyperparameters

Parameter| Value Per-device batch size| 1 Gradient accumulation| 8 Effective batch size| 8 Learning rate| 2e-4 Scheduler| Linear Warmup steps| 5 Weight decay| 0.01 Optimizer| AdamW 8-bit Training steps| 5 Precision| FP16 Max sequence length| 1536 tokens Maximum image dimension| 672 px Random seed| 3407

The NVIDIA T4 GPU does not support BF16, so FP16 was used.

Hardware and Training Cost

Property| Measurement Platform| Google Colab GPU| NVIDIA T4 GPU Memory| 16 GB Peak GPU Memory Usage| Approximately 8.4 GB Training Duration| Approximately 25 seconds Training Steps| 5

Logged training loss:

• Step 1: 4.79
• Step 5: 3.74

These loss values are not indicators of model quality because the training run was extremely short.

The exact training configuration is available in "training_config.json".

Training Data

The experimental dataset contains 36 examples in total.

Dataset Composition

24 Synthetic Image Examples

• OCR of short codes.
• Simple arithmetic represented in images.
• Counting colored circles.
• Examples written in Indonesian, English, and Japanese.

12 Handwritten Identity Examples

• Indonesian identity prompts.
• English identity prompts.
• Japanese identity prompts.
• Designed to teach the model to identify itself as a K-Ai model.

The synthetic dataset was generated primarily to validate the multimodal fine-tuning pipeline.

It is not a meaningful production training dataset.

Larger and more diverse real-world multimodal datasets, particularly for Indonesian and Japanese, are planned for future development.

Evaluation

No standard benchmarks have been conducted.

The model has not been evaluated using established benchmarks such as:

• MMMU
• DocVQA
• OCRBench
• MathVista
• Indonesian language benchmarks
• Japanese language benchmarks

No benchmark performance claims are made.

Pipeline Smoke Test

A small evaluation was conducted to verify that the inference pipeline works with and without the LoRA adapter.

The evaluation consisted of:

• 14 total items.
• 5 text questions.
• 9 synthetic-image questions.
• OCR, mathematics, and visual reasoning tasks.
• English, Indonesian, and Japanese.
• Greedy decoding.
• Exact expected-string containment scoring.

Results are available in "eval_results.json".

This evaluation is strictly a pipeline verification exercise. Its small size does not provide reliable evidence of model quality or general capabilities.

For published benchmark results of the underlying model, refer to the "Qwen3-VL-8B-Instruct model card" (https://huggingface.co/Qwen/Qwen3-VL-8B-Instruct).

Usage

The repository provides LoRA adapter weights. The original base model is required.

With Unsloth

from unsloth import FastVisionModel

model, processor = FastVisionModel.from_pretrained( "Keanualfahd/K-Oss-8-Mini-Instruct-Beta", load_in_4bit=True )

FastVisionModel.for_inference(model)

With Transformers and PEFT

from transformers import AutoModelForImageTextToText, AutoProcessor from peft import PeftModel

base = AutoModelForImageTextToText.from_pretrained( "Qwen/Qwen3-VL-8B-Instruct", torch_dtype="auto", device_map="auto" )

model = PeftModel.from_pretrained( base, "Keanualfahd/K-Oss-8-Mini-Instruct-Beta" )

processor = AutoProcessor.from_pretrained( "Qwen/Qwen3-VL-8B-Instruct" )

messages = [{ "role": "user", "content": [ { "type": "image", "image": "https://example.com/photo.jpg" }, { "type": "text", "text": "Jelaskan gambar ini." } ] }]

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

out = model.generate( **inputs, max_new_tokens=256 )

print( processor.decode( out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True ) )

Intended Use

This Beta release is intended for:

• Experimental research on compact vision-language models.
• Testing multilingual QLoRA fine-tuning workflows.
• Prototyping small AI assistants.
• Exploring multimodal capabilities across English, Indonesian, and Japanese.
• Early developer testing and feedback.
• Evaluating K-Ai model identity integration.

This release is not intended to represent the final capabilities of the K-Oss 8 Mini model family.

Limitations and Out-of-Scope Use

• Extremely small training run with only 5 optimization steps.
• Behavior is expected to remain close to the original base model.
• The training dataset contains only 36 examples.
• Language capabilities have not been independently benchmarked.
• OCR, mathematics, and visual reasoning may produce incorrect or fabricated answers.
• Not suitable for production deployment or high-stakes medical, legal, and financial decisions.
• Future datasets must have their licenses reviewed before commercial use.

Users should independently evaluate the model before integrating it into applications.

K-Ai Release Philosophy

K-Ai follows a flexible model release strategy.

Beta is an optional release stage, not a mandatory requirement for every model.

Depending on development progress, evaluation results, and release readiness, K-Ai models may follow different release paths.

Release Categories

Experimental

• Internal development.
• Early architecture testing.
• Training pipeline experiments.
• Preliminary evaluation.

Beta (Optional)

• Public testing.
• Additional validation.
• Community feedback.
• Identifying issues before a potential Production release.

Production

• Models that meet applicable quality and stability requirements.
• Suitable for their declared deployment purposes.
• Supported by appropriate evaluation and documentation.

A model may move directly from development to Production when it meets the required standards. Not every model needs a public Beta phase.

K-Ai releases when ready, not simply when a version number changes.

Attribution and License

• Built on "Qwen3-VL-8B-Instruct" (https://huggingface.co/Qwen/Qwen3-VL-8B-Instruct) by the Qwen team (Alibaba Cloud).
• The underlying base model is licensed under Apache-2.0.
• This adapter is released under Apache-2.0.
• The K-Oss 8 family relationship refers to a separate model based on NVIDIA Nemotron 3.5 Super VL. Refer to NVIDIA's model card for its respective licensing terms.

Contact

Developer: K-Ai (Kortex AI)

• Hugging Face Organization: https://huggingface.co/K-Ai-Official
• Developer Profile: https://huggingface.co/Keanualfahd