HiTZ/Latxa-Qwen3-VL-2B-Instruct

VISIONConcurrency Cost:1Model Size:2BQuant:BF16Ctx Length:32kPublished:Dec 12, 2025License:apache-2.0Architecture:Transformer Open Weights Cold

HiTZ/Latxa-Qwen3-VL-2B-Instruct is a 2 billion parameter multimodal and multilingual instruction-tuned model developed by HiTZ Research Center & IXA Research group, based on Qwen3-VL-2B-Instruct. This model is specifically adapted for improved performance in Basque, Galician, and Catalan, alongside English and Spanish, and excels at interactive instruction following with both text and image inputs. It is designed for use with low-resource languages, demonstrating significant performance gains on Basque-specific benchmarks.

Loading preview...

Overview

Latxa-Qwen3-VL-2B-Instruct is a 2 billion parameter vision-language model developed by the HiTZ Research Center & IXA Research group, building upon the Qwen3-VL-2B-Instruct architecture. Its primary innovation lies in its adaptation for low-resource languages, particularly Basque, Galician, and Catalan, while also supporting Spanish and English. The model is capable of understanding and generating text, as well as processing images, making it a multimodal instruct model.

Key Capabilities

  • Multimodal Instruction Following: Processes both text and image inputs to follow instructions and act as a chat assistant.
  • Low-Resource Language Adaptation: Specifically fine-tuned for enhanced performance in Basque, Galician, and Catalan, addressing challenges in languages with limited digital resources.
  • Multilingual Support: Available in a multi variant adapted for Basque, Galician, and Catalan, and a mono_eu variant adapted solely for Basque.
  • Improved Performance: Benchmarks show significant performance increases on Basque tasks (e.g., +14.02% average gain on EU tasks for the multi variant) compared to the base Qwen3-VL 2B model.

Good For

  • Applications requiring Basque, Galician, or Catalan language processing: Ideal for tasks and applications targeting these specific low-resource languages.
  • Multimodal tasks in supported languages: Suitable for scenarios where both image understanding and text generation/instruction following are needed.
  • Research in low-resource NLP: Provides a strong baseline and adapted model for further research into language models for less-resourced languages, as detailed in the paper "Instructing Large Language Models for Low-Resource Languages: A Systematic Study for Basque" (link to paper).

Finetunes
1 models