HiTZ/Latxa-Qwen3-VL-32B-Instruct

VISIONConcurrency Cost:2Model Size:33.4BQuant:FP8Ctx Length:32kPublished:Feb 19, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Cold

HiTZ/Latxa-Qwen3-VL-32B-Instruct is a 33.4 billion parameter vision-language instruct model developed by HiTZ Research Center & IXA Research group. Built upon Qwen3-VL-32B-Instruct, this model is specifically adapted for enhanced performance in Basque, Galician, and Catalan, alongside its existing multilingual capabilities. It excels at understanding and generating text from images, particularly for instruction-following tasks in low-resource languages.

Loading preview...

Model Overview

HiTZ/Latxa-Qwen3-VL-32B-Instruct is a 33.4 billion parameter Vision-Language Instruct Model developed by the HiTZ Research Center & IXA Research group. It is an adaptation of the powerful Qwen3-VL-32B-Instruct, specifically fine-tuned to improve performance in Basque, Galician, and Catalan, while also supporting other languages like Spanish and English.

Key Capabilities

  • Multimodal Understanding: Processes both text and image inputs to generate relevant text outputs.
  • Multilingual Adaptation: Features distinct revisions for language focus:
    • multi variant: Adapted for Basque, Galician, and Catalan.
    • mono_eu variant: Adapted specifically for Basque.
  • Instruction Following: Designed to follow instructions and function as a chat assistant.
  • Low-Resource Language Focus: Demonstrates significant performance improvements on Basque tasks compared to its base model, with average gains of over 15% on various benchmarks.

Intended Use Cases

  • Basque Language Applications: Optimized for use with Basque data, offering improved accuracy and fluency.
  • Multilingual Scenarios: The multi variant is suitable for applications involving Basque, Galician, and Catalan.
  • Instruction-based Interactions: Ideal for chatbots and systems requiring instruction following in the specified languages.

Limitations

  • Performance is not guaranteed for languages other than those it was specifically adapted for.
  • As a derivative of Qwen3-VL, it may inherit similar biases, risks, and limitations.