HiTZ/Latxa-Qwen3.5-4B

VISIONConcurrent Unit Cost:1Model Size:4.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jun 24, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

HiTZ/Latxa-Qwen3.5-4B is a 4.5 billion parameter multimodal and multilingual instruct model developed by HiTZ Research Center & IXA Research group, adapted from Qwen3.5-4B. It is specifically fine-tuned for improved performance in Basque, Galician, and Catalan, while also supporting English and Spanish. This vision-language model excels at understanding and generating text, processing images, and following interactive instructions, particularly for low-resource languages.

Loading preview...

Overview

HiTZ/Latxa-Qwen3.5-4B is a 4.5 billion parameter Vision-Language Instruct Model developed by the HiTZ Research Center & IXA Research group. It is built upon Qwen3.5-4B and has been specifically adapted for enhanced performance in Basque, Galician, and Catalan, in addition to supporting English and Spanish. This model is capable of understanding and generating text, processing images, and following interactive instructions.

Key Capabilities

  • Multimodal Understanding: Processes both text and image inputs, enabling rich interactive experiences.
  • Multilingual Adaptation: Significantly improved performance for Basque, Galician, and Catalan, making it suitable for low-resource language applications.
  • Instruction Following: Trained to follow instructions and function as a chat assistant.
  • Enhanced Performance: Demonstrates substantial gains across various Basque-specific benchmarks (e.g., +26.81 on Eus Proficiency, +22.08 on BertaQA local) compared to the base Qwen3.5-4B model.

Good for

  • Applications requiring multimodal understanding (image and text) in Basque, Galician, or Catalan.
  • Developing chat assistants or instruction-following agents for these specific low-resource languages.
  • Research and development in multilingual LLMs, particularly for languages with limited digital resources.

Limitations

  • While adapted for multiple languages, optimal performance is guaranteed primarily for Basque, Galician, and Catalan. Performance for other languages is not explicitly guaranteed.
  • As it is based on Qwen3.5 models, it may inherit similar biases, risks, and limitations. Users should conduct proper risk assessment for production environments.

Finetunes
1 models