K-Oss 8 Mini Instruct Documentation
Official Technical Documentation by K-Ai
Welcome to the official documentation for K-Oss 8 Mini Instruct, a multilingual instruction-following model developed by K-Ai (Kortex AI).
This documentation provides information about the model architecture, training dataset, supported languages, inference, evaluation, and deployment.
- Introduction
K-Oss 8 Mini Instruct is a fine-tuned Vision-Language Model based on Qwen3-VL-8B.
Developed by K-Ai, the model focuses on multilingual conversational capabilities and instruction following, with support for Indonesian, English, and Japanese.
The project aims to explore accessible multilingual AI development and practical applications across different languages.
- Model Information
Property| Details Model Name| K-Oss 8 Mini Instruct Developer| K-Ai Base Model| Qwen3-VL-8B Architecture| Qwen3-VL Model Type| Fine-tuned Instruct Training Dataset| "kai_synth_v1" License| Apache 2.0
- Architecture
K-Oss 8 Mini Instruct is built upon the Qwen3-VL-8B architecture.
The base architecture provides the foundation for language and vision-language processing.
K-Ai applies additional fine-tuning to adapt the model toward its intended instruction-following and multilingual use cases.
The current release uses a LoRA adapter rather than a complete set of independently distributed model weights.
- Training Dataset
Dataset Name: "kai_synth_v1"
The "kai_synth_v1" dataset is used in the fine-tuning process of K-Oss 8 Mini Instruct.
Its intended role is to support the model's instruction-following behavior and multilingual adaptation.
Primary Languages
- Indonesian
- English
- Japanese
Detailed dataset composition, collection methodology, and training configuration may be published in future documentation updates.
- Language Support
K-Oss 8 Mini Instruct is designed to support three primary languages.
Language| Code Indonesian| "id" English| "en" Japanese| "ja"
The model aims to provide responses consistent with the language used in user instructions.
English may serve as a fallback language when necessary.
- Intended Applications
K-Oss 8 Mini Instruct is intended for:
- General-purpose AI conversations.
- Multilingual question answering.
- Educational assistance.
- Text generation and summarization.
- Instruction-following applications.
- Developer experimentation.
- Multimodal research, subject to validated capabilities.
- Inference
K-Oss 8 Mini Instruct can be integrated into compatible inference environments.
Developers should follow the model repository's recommended loading procedure and use the appropriate tokenizer, processor, and chat template.
Since the current release uses a LoRA adapter, developers must load it alongside its compatible base model.
For vision-language inputs, a compatible image processor and supported multimodal input pipeline are required.
- Evaluation and Benchmarks
K-Ai evaluates K-Oss 8 Mini Instruct across several dimensions:
- In-domain instruction following.
- General question answering.
- Multilingual response consistency.
- Language accuracy.
- Comparison with the original base model.
K-Ai aims to publish reproducible evaluation results as testing progresses.
Benchmark results should be interpreted according to the evaluation methodology and test dataset used.
The model should not be used as the sole authority for high-stakes decisions.
- License
K-Oss 8 Mini Instruct is distributed under the Apache License 2.0, subject to applicable base-model licensing and redistribution requirements.
Full license text:
https://www.apache.org/licenses/LICENSE-2.0
Users should review the original Qwen3-VL-8B license and applicable attribution requirements before redistribution.
- Acknowledgements
K-Ai acknowledges the Qwen team and contributors for developing Qwen3-VL-8B, which serves as the foundation for K-Oss 8 Mini Instruct.
K-Oss is an independent K-Ai model adaptation and is not an official Qwen release.
K-Ai — Kortex AI
We build because we believe we can.