Alibaba's Ovis 2.5: Open-Source Multimodal LLM Breakthrough
Alibaba’s AIDC-AI team has unveiled Ovis 2.5, its latest large multimodal language model (MLLM), making a significant splash in the open-source artificial intelligence community. Available in 9-billion and 2-billion parameter versions, Ovis 2.5 introduces pivotal technical advancements that redefine performance and efficiency benchmarks for MLLMs, particularly in handling high-detail visual information and complex reasoning tasks that have long challenged the field.
A cornerstone of Ovis 2.5’s innovation lies in its native-resolution vision transformer (NaViT). This allows the model to process images at their original, varying resolutions, a stark departure from previous approaches that often relied on tiling or forced resizing. Such older methods frequently resulted in the loss of vital global context and intricate details. By preserving the full integrity of both complex charts and natural images, NaViT enables Ovis 2.5 to excel at visually dense tasks, from interpreting scientific diagrams to analyzing elaborate infographics and forms.
Beyond enhanced visual perception, Ovis 2.5 tackles the intricacies of reasoning with a sophisticated training curriculum. This goes beyond standard chain-of-thought supervision by incorporating “thinking-style” samples designed for self-correction and reflection. The culmination of this approach is an optional “thinking mode” at inference time. While enabling this mode may trade some response speed, it significantly boosts step-by-step accuracy and allows for deeper model introspection, proving particularly advantageous for tasks demanding profound multimodal analysis, such as scientific question answering or intricate mathematical problem-solving.
Ovis 2.5’s capabilities are reflected in its impressive benchmark results. The larger Ovis 2.5-9B model achieved an average score of 78.3 on the OpenCompass multimodal leaderboard, positioning it as a leading contender among all open-source MLLMs under 40 billion parameters. Its more compact sibling, Ovis 2.5-2B, also set a new standard for lightweight models, scoring 73.9, making it an ideal candidate for on-device or resource-constrained applications. Both models demonstrate exceptional performance across specialized domains, outperforming open-source competitors in areas like STEM reasoning (validated on datasets such as MathVista, MMMU, and WeMath), optical character recognition (OCR) and chart analysis (as seen on OCRBench v2 and ChartQA Pro), visual grounding (RefCOCO, RefCOCOg), and comprehensive video and multi-image understanding (BLINK, VideoMME). Online discussions among AI developers have particularly lauded the advancements in OCR and document processing, highlighting the model’s improved ability to extract text from cluttered images, understand complex forms, and handle diverse visual queries with flexibility.
Efficiency is another hallmark of Ovis 2.5. The models optimize end-to-end training through techniques like multimodal data packing and advanced hybrid parallelism, yielding up to a threefold or even fourfold speedup in overall throughput. Furthermore, the lightweight 2-billion parameter variant embodies a “small model, big performance” philosophy, extending high-quality multimodal understanding to mobile hardware and edge devices, thereby democratizing access to advanced AI capabilities.
Alibaba’s Ovis 2.5 models represent a significant stride forward in open-source multimodal AI. By integrating a native-resolution vision transformer and an innovative “thinking mode” for deeper reasoning, Ovis 2.5 not only achieves state-of-the-art results across critical benchmarks but also narrows the performance gap with proprietary AI solutions. Its focus on efficiency and accessibility ensures that advanced multimodal understanding is within reach for both cutting-edge researchers and practical, resource-constrained applications.