Language support
English, Indonesian and Japanese are the target languages. Each was trained on and tested separately (passed/total on small held-out sets):
Table with columns: Language, In-domain tests (fine-tuned), In-domain tests (base), General Q&A (fine-tuned), Replies in the right language| Language | In-domain tests (fine-tuned) | In-domain tests (base) | General Q&A (fine-tuned) | Replies in the right language |
|---|
| Indonesian | 17/20 | 15/20 | 16/20 | 19/20 |
| English | 17/20 | 17/20 | 17/20 | 19/20 |
| Japanese | 14/20 | 12/20 | 13/20 | 16/20 |
Model details
Table | |
|---|
| Developer | K-Ai (Kortex AI) |
| Base model | Qwen/Qwen3-VL-8B-Instruct (about 8.8B parameters) |
| Fine-tuning | QLoRA, r=16, alpha=16 |
| Input / output | Text and images in, text out |
| Languages | English, Indonesian, Japanese |
| License | Apache-2.0 (follows the base model) |
Training procedure
- Method: supervised fine-tuning (TRL
SFTTrainer, Unsloth vision collator) on 4-bit quantized base weights, then merged to 16-bit
- Trainable: attention and MLP modules of the language model; frozen: vision tower
- Hyperparameters: [[steps]] steps, batch size 1 x gradient accumulation 8 (effective 8), learning rate [[1e-4]], AdamW 8-bit, weight decay 0.01, fp16, max sequence length 1536, images resized to at most 640 px, seed 3407
- Hardware: Google Colab, one NVIDIA T4 (16 GB); training time [[minutes]]; loss [[start]] -> [[end]]
Training data
[[total]] examples, mostly synthetic (dataset kai-synth-v1, generated by a script with exactly computed answers), in all three languages:
- Images: OCR (codes, numbers, words), receipts (Rp / $ / ¥), bar charts, tables, forms, math in images, shape counting
- Text: word problems, general facts, translation between the three languages, identity, and a small set of refusal / benign security examples
- Held out for testing: the last question template per task and language, 20% of OCR words, selected facts and phrases
- Some text examples (identity, translation phrases, facts, refusal and benign answers) were written with AI assistance and were not reviewed by native speakers.
Usage
from transformers import AutoModelForImageTextToText, AutoProcessor model = AutoModelForImageTextToText.from_pretrained( "Keanualfahd/K-Oss-8-Mini-Instruct", torch_dtype="auto", device_map="auto")processor = AutoProcessor.from_pretrained("Keanualfahd/K-Oss-8-Mini-Instruct") messages = [{"role": "user", "content": [ {"type": "image", "image": "https://example.com/receipt.jpg"}, {"type": "text", "text": "Berapa total pada struk 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=128)print(processor.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))
Intended use
K-Oss 8 Mini Instruct is a multilingual instruction-following model developed by K-Ai, based on Qwen3-VL-8B.
Primary Use Cases
- General-purpose conversational AI.
- Indonesian, English, and Japanese language assistance.
- Question answering and information assistance.
- Text generation, summarization, and basic reasoning.
- Educational assistance and learning support.
- Multimodal interaction, subject to the capabilities retained from the base architecture.
Target Users
- Developers integrating AI into applications.
- Students and educators.
- AI researchers and enthusiasts.
- Users exploring multilingual AI systems.
Out-of-Scope Uses
- Medical, legal, or financial decisions without qualified human oversight.
- Generating harmful, deceptive, or misleading content.
- High-stakes automated decision-making.
This model is intended for research, experimentation, and general-purpose AI applications. Its outputs should be reviewed for accuracy and reliability.
License and attribution
Apache-2.0, following Qwen3-VL-8B-Instruct by the Qwen team (Alibaba Cloud). K-Oss 8's own license terms follow its Nemotron base and are separate. Fine-tuned with Unsloth and TRL.
Ringkasan (Bahasa Indonesia)
K-Oss 8 Mini Instruct adalah model multimodal kecil dari keluarga K-Oss 8 buatan K-Ai. Model ini membaca teks dan gambar dan menjawab dalam bahasa Inggris, Indonesia, dan Jepang. Hasil fine-tuning dari Qwen3-VL-8B-Instruct, sebagian besar dilatih dengan data sintetis. Status: Public.
概要(日本語)
K-Oss 8 Mini Instructは、K-Ai社のK-Oss 8シリーズの小型マルチモーダルモデルです。テキストと画像を理解し、英語、インドネシア語、日本語で応答します。こちらはyバージョンです。