HiTZ/Latxa-Qwen3-VL-4B-Instruct

VISIONConcurrency Cost:1Model Size:4BQuant:BF16Ctx Length:32kPublished:Dec 12, 2025License:apache-2.0Architecture:Transformer0.0K Open Weights Cold

Latxa-Qwen3-VL-4B-Instruct is a 4 billion parameter vision-language instruct model developed by HiTZ Research Center and IXA Research group. Built on Qwen3-VL-4B-Instruct, it is specifically adapted for improved performance in Basque, Galician, and Catalan, alongside Spanish and English. This model excels at understanding and generating text from image inputs, making it suitable for multimodal instruction following in low-resource languages.

Loading preview...

Overview

Latxa-Qwen3-VL-4B-Instruct is a 4 billion parameter multimodal and multilingual instruct model developed by the HiTZ Research Center and IXA Research group. It is based on Qwen3-VL-4B-Instruct and has been specifically adapted to enhance performance in Basque, Galician, and Catalan, in addition to supporting Spanish and English. The model is available in two main revisions: a multi variant adapted for Basque, Galician, and Catalan, and a mono_eu variant adapted solely for Basque.

Key Capabilities

  • Multimodal Understanding: Processes both image and text inputs to generate relevant text outputs.
  • Multilingual Adaptation: Significantly improved performance for low-resource languages like Basque, Galician, and Catalan, as demonstrated by evaluation scores.
  • Instruction Following: Designed to follow instructions and function as a chat assistant.
  • Performance Gains: Evaluation shows substantial improvements over the base Qwen3-VL 4B model across various Basque tasks, with average gains of +14.78% for mono_eu and +15.86% for multi variants.

Good For

  • Basque Language Applications: Ideal for use cases requiring strong performance in Basque, including text generation, question answering, and instruction following.
  • Multilingual Low-Resource Scenarios: Suitable for applications involving Galician and Catalan, leveraging its specialized adaptation.
  • Vision-Language Tasks: Effective for scenarios where understanding and responding to image content is crucial, such as image captioning or visual question answering in supported languages.

Finetunes
1 models