ATH-MaaS/Ovis

A novel Multimodal Large Language Model (MLLM) architecture, designed to structurally align visual and textual embeddings.

What it solves

Ovis is a Multimodal Large Language Model (MLLM) designed to bridge the gap between visual and textual embeddings. It addresses the challenge of high-resolution image processing and complex reasoning in multimodal tasks, providing a way to structurally align visual data with language model understanding.

How it works

Ovis uses a novel architecture that aligns visual embeddings from a Vision Transformer (ViT) with the embeddings of a Large Language Model (LLM). It supports various LLM backbones (such as Qwen3) and ViT backbones (such as SigLIP2). The latest version, Ovis2.5, introduces a "thinking mode" for reflective reasoning, allowing the model to process images at native resolution and handle multi-image or video inputs.

Who it’s for

This project is for AI researchers and developers who need a high-performance open-source MLLM capable of handling STEM tasks, chart analysis, grounding, and video understanding.

Highlights

  • Reflective Reasoning: A dedicated "thinking mode" that enhances the model's ability to perform complex reasoning tasks.
  • Native-Resolution Perception: Ability to process visual inputs at their original resolution for better detail preservation.
  • Flexible Architecture: Compatible with multiple LLM and ViT backbones, with models ranging from 2B to 34B parameters.
  • Broad Modality Support: Handles single images, multiple images, video, and pure text inputs.
  • Fine-tuning Support: Includes built-in training scripts and compatibility with the ms-swift framework.

Related

  • Project
    ATH-MaaS/Ovis-Image

    Ovis-Image is a 7B text-to-image model optimized for high-quality text rendering and layout precision, designed to run efficiently on accessible hardware.

  • Project
    zai-org/GLM-V

    A series of vision-language models (VLMs) that enhance multimodal reasoning, supporting native function calling, visual grounding, and complex document understanding.

  • Project
    inclusionAI/Ming

    Ming-flash-omni 2.0 is an open-source omni-MLLM that unifies multimodal perception and generation across text, image, audio, and video in a single MoE-based architecture.

  • Dispatch
    OpenAI o3 and o4-mini Visual Reasoning Release

    OpenAI has introduced o3 and o4-mini, the first models in the o-series capable of incorporating image manipulation tools directly into their internal chain-of-thought for advanced visual reasoning.

  • Dispatch
    Hugging Face Vision Language Models 2025 Update

    Hugging Face provides a comprehensive overview of the 2024-2025 evolution of Vision Language Models (VLMs), highlighting trends in any-to-any architectures, reasoning models, and the rise of Vision-Language-Action (VLA) models for robotics.