sengtha/Qwen3-0.6B-khm-ft3
The sengtha/Qwen3-0.6B-khm-ft3 model is a 0.6 billion parameter language model, fine-tuned from Qwen3-0.6B by sengtha, specifically optimized for Khmer language processing. It features a trimmed 32k Khmer-focused vocabulary to enable efficient on-device execution, even on older smartphones. This model excels at grounded Khmer question-answering and Retrieval Augmented Generation (RAG) for offline use, making it suitable for applications requiring local language support on resource-constrained devices.
Loading preview...
sengtha/Qwen3-0.6B-khm-ft3: On-Device Khmer RAG Model
This model is a 0.6 billion parameter language model, developed by sengtha, specifically engineered for Khmer language processing on resource-constrained devices. Fine-tuned from Qwen3-0.6B, its primary differentiator is a trimmed 32k Khmer-focused vocabulary from the original ~150k Qwen3 vocabulary. This reduction significantly shrinks the output logits buffer, allowing the model to run efficiently on devices with limited RAM, such as a 2019 Galaxy S10, and enabling offline functionality.
Key Capabilities
- Optimized for Khmer: Continued pre-training on FineWeb-2 (Khmer) and ParaCrawl, followed by instruction tuning on
sengtha/khmer-qa, ensures strong Khmer fluency and accurate, fuller answers. - On-Device Performance: The specialized vocabulary and small parameter count facilitate deployment on mobile devices, making it ideal for applications requiring local processing.
- Grounded Q&A / RAG: Designed for question-answering where context is retrieved from user documents, providing relevant answers in Khmer.
- Qwen3 Chat Template: Utilizes the standard Qwen3 chat template (ChatML) for interaction, with an optional
/no_thinkappend for faster, direct responses.
Intended Use Cases
- Offline Khmer RAG applications: Particularly suited for scenarios where internet connectivity is unreliable or privacy concerns necessitate local data processing.
- Mobile-first language solutions: Enables rich Khmer language interaction directly on smartphones and other edge devices.
- Khmer question-answering systems: Provides a specialized tool for generating responses to Khmer queries based on provided context.