We present Ovis2.5, a successor to Ovis2 designed for native-resolution visual perception and strong multimodal reasoning. Ovis2.5 integrates a native-resolution vision transformer that processes images at their native, variable resolutions, avoiding the degradation from fixed-resolution tiling and preserving both fine detail and global layout -- crucial for visually dense content like complex charts. To strengthen reasoning, we train the model to move beyond linear chain-of-thought and perform reflection -- including self-checking and revision. This advanced capability is exposed as an optional "thinking mode" at inference time, allowing users to trade latency for enhanced accuracy on difficult inputs. The model is trained via a comprehensive five-phase curriculum that progressively builds its skills. The process begins with foundational visual and multimodal pretraining, advances through large-scale instruction tuning, and culminates in alignment and reasoning enhancement using DPO and GRPO. To scale these upgrades efficiently, we employ multimodal data packing and hybrid parallelism, yielding a significant end-to-end speedup. We release two open-source models: Ovis2.5-9B and Ovis2.5-2B. The latter continues the "small model, big performance" philosophy of Ovis2, making it ideal for resource-constrained, on-device scenarios. On the OpenCompass multimodal leaderboard, Ovis2.5-9B averages 78.3, marking a substantial improvement over its predecessor, Ovis2-8B, and achieving state-of-the-art results among open-source MLLMs in the sub-40B parameter range; Ovis2.5-2B scores 73.9, establishing SOTA for its size. Beyond aggregate scores, Ovis2.5 achieves leading results on STEM benchmarks, exhibits strong capabilities on grounding and video tasks, and achieves open-source SOTA at its scale for complex chart analysis.
核心贡献 · Key contributions
集成原生分辨率视觉 Transformer（NaViT），以可变原生分辨率处理图像，避免固定分辨率分块，保留细节和全局布局。 Integrates a native-resolution vision transformer (NaViT) to process images at variable native resolutions, avoiding fixed-resolution tiling and preserving fine details and global layout.
在推理时引入可选的“思考模式”，支持反思、自我检查和修正，实现超越线性思维链的深度推理。 Introduces an optional 'thinking mode' at inference time, enabling reflection, self-checking, and revision for deeper reasoning beyond linear chain-of-thought.
在 OpenCompass 多模态排行榜上，Ovis2.5-9B 以 78.3 分在 40B 参数以下开源 MLLM 中取得最优结果。 Achieves state-of-the-art results on OpenCompass multimodal leaderboard among open-source MLLMs under 40B parameters, with Ovis2.5-9B scoring 78.3.
在 STEM、图表分析、视觉定位和视频基准上取得领先性能，包括复杂图表分析的开源最优。 Delivers leading performance on STEM, chart analysis, visual grounding, and video benchmarks, including open-source SOTA on complex chart analysis.
采用五阶段课程训练流程，结合多模态数据打包和混合并行，实现 3-4 倍端到端加速。 Employs a five-phase curriculum training pipeline with multimodal data packing and hybrid parallelism, achieving 3-4x end-to-end speedup.
发布两个开源模型 Ovis2.5-9B 和 Ovis2.5-2B，后者在 OpenCompass 上以 73.9 分取得同尺寸最优。 Releases two open-source models, Ovis2.5-9B and Ovis2.5-2B, with the latter achieving SOTA for its size on OpenCompass (73.9).
局限 · Limitations
将感知扩展到 4K 级高分辨率图像同时保持准确性仍是一个开放挑战。 Scaling perception to 4K-level high-resolution images while maintaining accuracy remains an open challenge.
处理长输入视频并实现更丰富的时间推理尚未完全解决。 Handling long-input video with richer temporal reasoning is not fully addressed.
更紧密地集成工具使用以增强行动推理留待未来工作。 Tighter integration of tool use for action-augmented reasoning is left for future work.
“思考模式”以延迟换取准确性，可能不适合对延迟敏感的应用。 The 'thinking mode' trades latency for accuracy, which may not suit latency-sensitive applications.
评估限于基准测试；实际部署性能可能有所不同。 Evaluation is limited to benchmarks; real-world deployment performance may vary.