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

K-Oss 8 Mini Instruct

K-Oss 8 Mini Instruct is the compact, instruction-tuned member of the K-Oss 8 family by K-Ai (Kortex AI). It supports English, Indonesian, and Japanese (日本語), and accepts text and images.

This repository contains a LoRA adapter and requires the base model Qwen/Qwen3-VL-8B-Instruct.

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 (this repository): the small version of the family. It is a LoRA adapter (about 43.6M trainable parameters) on top of the 8.8B-parameter Qwen3-VL-8B-Instruct.

Important: K-Oss 8 Mini Instruct is a separate, smaller model. It is built on Qwen3-VL-8B-Instruct. It does not use Nemotron weights and is not distilled from K-Oss 8. It shares the family name, goals, and language focus. Do not assume it matches the capabilities of the 120B model.

Supported languages

English, Indonesian, and Japanese are the target languages. They are goals of the fine-tune and have not been benchmarked.

Base model attribution

• Built on Qwen3-VL-8B-Instruct by the Qwen team (Alibaba Cloud), Apache-2.0.
• The K-Oss 8 family relationship refers to a separate model based on NVIDIA Nemotron 3.5 Super VL. See NVIDIA's model card for that model's own license terms.

Training methodology

• QLoRA (4-bit) with Unsloth, LoRA r=16, alpha=16
• Attention and MLP layers of the language model trained; vision tower frozen
• Hardware: Google Colab, NVIDIA T4 (16 GB), fp16

Release status

This release is an early pipeline smoke test: 5 training steps on 36 examples (24 synthetic image examples and 12 identity examples). It mainly verifies the training pipeline and adds a K-Ai identity. Expect behavior close to the base model.

Intended capabilities

Inherited from the base model: image understanding, visual question answering, OCR, math and science reasoning, and general instruction following. These capabilities have not been verified beyond a very small smoke test (see eval_results.json in this repository). It is not a benchmark.

Usage

from unsloth import FastVisionModel

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

Limitations

• Very small training run; quality depends on how much real multimodal data is used in future versions.
• May produce incorrect or fabricated answers, including in OCR and math.
• Language support is a target and not measured.
• Not suitable for medical, legal, or financial decisions.
• Dataset licenses (some data are GPT-4 generated) should be reviewed before commercial use.

License

This adapter is released under Apache-2.0, following the Qwen3-VL-8B-Instruct base model. Review dataset licenses separately.