xCloudinfo/Gemma-4-26B-A4B-TAIDE-zhTW

VISIONConcurrent Unit Cost:2Model Size:26BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jun 8, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

xCloudinfo/Gemma-4-26B-A4B-TAIDE-zhTW is a 26 billion parameter multimodal language model developed by xCloudinfo, based on Google's Gemma-4-26B-A4B-it architecture. This model is specifically fine-tuned with TAIDE-distilled traditional Chinese (Taiwan) self-instruct data, retaining its native image recognition capabilities. It excels in image-text-to-text tasks and conversational AI in traditional Chinese, making it suitable for applications requiring both visual understanding and nuanced Taiwanese Mandarin interaction.

Loading preview...

Overview

xCloudinfo/Gemma-4-26B-A4B-TAIDE-zhTW is a 26 billion parameter multimodal model, developed by xCloudinfo, that specializes in traditional Chinese (Taiwan) language processing. It is built upon Google's gemma-4-26B-A4B-it base model, which features a 128-expert MoE architecture and supports both text and vision modalities. The model has been fine-tuned using bf16 LoRA SFT with a cleaned dataset of 27,955 traditional Chinese self-instruct entries, distilled from TAIDE models.

Key Capabilities

  • Multimodal Understanding: Retains the base model's native image recognition capabilities, allowing for image-text-to-text generation. The LoRA fine-tuning specifically targeted the language model, leaving the vision tower unchanged.
  • Traditional Chinese (Taiwan) Optimization: Specifically trained on TAIDE-derived data, making it highly proficient in Taiwanese Mandarin nuances and context.
  • Efficient Fine-tuning: Utilizes bf16 LoRA for fine-tuning, which was then merged back into the full model.

Good For

  • Applications requiring multimodal conversational AI in traditional Chinese, particularly for the Taiwanese market.
  • Use cases involving image-text understanding and generation where traditional Chinese output is critical.
  • Developers looking for a robust model with integrated vision capabilities and strong performance in a specific regional language variant.