🎯 Introduction

Youtu-VL is a lightweight yet robust Vision-Language Model (VLM) built on the Youtu-LLM with 4B parameters. It pioneers Vision-Language Unified Autoregressive Supervision (VLUAS), which markedly strengthens visual perception and multimodal understanding. This enables a standard VLM to perform vision-centric tasks without task-specific additions. Across benchmarks, Youtu-VL stands out for its versatility, achieving competitive results on both vision-centric and general multimodal tasks.

✨ Key Features

• Comprehensive Vision-Centric Capabilities: The model demonstrates strong, broad proficiency across classic vision-centric tasks, delivering competitive performance in visual grounding, image classification, object detection, referring segmentation, semantic segmentation, depth estimation, object counting, and human pose estimation.

• Promising Performance with High Efficiency: Despite its compact 4B-parameter architecture, the model achieves competitive results across a wide range of general multimodal tasks, including general visual question answering (VQA), multimodal reasoning and mathematics, optical character recognition (OCR), multi-image and real-world understanding, hallucination evaluation, and GUI agent tasks.

🤗 Model Download

🧠 Model Architecture Highlights

• Vision–Language Unified Autoregressive Supervision (VLUAS): Youtu-VL is built on the VLUAS paradigm to mitigate the text-dominant optimization bias in conventional VLMs, where visual signals are treated as passive conditions and fine-grained details are often dropped. Rather than using vision features only as inputs, Youtu-VL expands the text lexicon into a unified multimodal vocabulary through a learned visual codebook, turning visual signals into autoregressive supervision targets. Jointly reconstructing visual tokens and text explicitly preserves dense visual information while strengthening multimodal semantic understanding.

• Vision-Centric Prediction with a Standard Architecture (no task-specific modules): Youtu-VL treats image and text tokens with equivalent autoregressive status, empowering it to perform vision-centric tasks for both dense vision prediction (e.g., segmentation, depth) and text-based prediction (e.g., grounding, detection) within a standard VLM architecture, eliminating the need for task-specific additions. This design yields a versitile general-purpose VLM, allowing a single model to flexibly accommodate a wide range of vision-centric and vsion-language requirements.

🏆 Model Performance

Vision-Centric Tasks

General Multimodal Tasks

🚀 Quickstart

This guide will help you quickly deploy and invoke the Youtu-VL-4B-Instruct-GGUF model using llama.cpp.

llama-server -hf tencent/Youtu-VL-4B-Instruct-GGUF:Q8_0 \
  --port 8080 \
  --image-max-tokens 2048 \
  --temp 0.1 \
  --top-p 0.001 \
  --repeat-penalty 1.05 \
  -n 12280 \
  --host 0.0.0.0

⚠️ Note: Dense prediction tasks (including Segmentation and Depth Estimation) are currently NOT supported in the Youtu-VL-4B-Instruct-GGUF version. For these capabilities, please refer to the original Transformers version: Youtu-VL-4B-Instruct.

🎉 Citation

If you find our work useful in your research, please consider citing our paper:

@article{youtu-vl,
  title={Youtu-VL: Unleashing Visual Potential via Unified Vision-Language Supervision},
  author={Tencent Youtu Lab},
  year={2026},
  eprint={2601.19798},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2601.19798}, 
}

@article{youtu-llm,
  title={Youtu-LLM: Unlocking the Native Agentic Potential for Lightweight Large Language Models},
  author={Tencent Youtu Lab},
  year={2025},
  eprint={2512.24618},
  archivePrefix={arXiv},
  primaryClass={cs.CL},
  url={https://arxiv.org/abs/2512.24618}, 
}