HiTZ/Latxa-Qwen3-VL-8B-Instruct
HiTZ/Latxa-Qwen3-VL-8B-Instruct is an 8 billion parameter multimodal and multilingual instruction-tuned vision-language model developed by HiTZ Research Center & IXA Research group. Built upon Qwen3-VL-8B-Instruct, it is specifically adapted for enhanced performance in Basque, Galician, and Catalan, alongside English and Spanish. This model excels at understanding and generating text from image inputs, making it suitable for instruction-following and chat assistance in low-resource languages.
Loading preview...
Model Overview
HiTZ/Latxa-Qwen3-VL-8B-Instruct is an 8 billion parameter vision-language instruct model developed by the HiTZ Research Center & IXA Research group. It is an adaptation of the powerful Qwen3-VL-8B-Instruct, specifically fine-tuned to improve performance in low-resource languages, particularly Basque, Galician, and Catalan, while also supporting Spanish and English.
Key Capabilities
- Multimodal Understanding: Processes both text and image inputs to generate responses.
- Multilingual Adaptation: Features
multiandmono_euvariants, with themultivariant adapted for Basque, Galician, and Catalan, andmono_eufor Basque only. - Instruction Following: Designed to act as a chat assistant and follow complex instructions.
- Improved Low-Resource Language Performance: Demonstrates significant performance gains over the base Qwen3-VL model on Basque-specific benchmarks, with average improvements of +15.86% for the
multivariant on Basque tasks.
Use Cases
- Basque Language Applications: Primarily intended for use with Basque data, offering enhanced performance in this language.
- Multilingual Chatbots: The
multivariant can be used for instruction-following and chat assistance in Basque, Galician, and Catalan. - Vision-Language Tasks: Suitable for scenarios requiring understanding of images combined with text-based instructions or queries in the supported languages.
Limitations
- Performance is not guaranteed for languages other than those it was specifically adapted for (Basque, Galician, Catalan).
- The models are still under development, and results for other languages will be released in the future.