Ovis2.5 Technical Report

1 Introduction

• Native-resolution NaViT processes images at variable native resolutions, avoiding lossy tiling and preserving fine detail and global layout for dense visuals.This addresses fixed-resolution front ends that can compromise global structure, a limitation identified in Ovis2.Source: §1 · p. 2
• Reflective-reasoning training teaches self-checking and revision beyond linear CoT, with optional thinking mode trading inference latency for accuracy on difficult inputs.Ovis2’s linear-CoT-focused training lacked reflective, self-corrective supervision for deeper reasoning.Source: §1 · p. 2
• A five-phase curriculum progresses from visual and multimodal pretraining through instruction tuning, DPO, and GRPO, while data packing and hybrid parallelism yield 3–4× end-to-end speedup.The phases are visual pretraining, multimodal pretraining, instruction tuning, DPO, and reinforcement learning with GRPO.Source: §1 · p. 2
• Ovis2.5-9B scores 78.3 on OpenCompass, improving substantially over Ovis2-8B and setting sub-40B open-source MLLM SOTA; Ovis2.5-2B scores 73.9, SOTA among comparable-size models.The 2B release continues the series’ small-model, big-performance approach for resource-constrained and on-device use.Source: §1 · p. 2
• Beyond aggregate scores, Ovis2.5 leads similar-scale open-source models across STEM benchmarks, performs strongly on grounding and video, and sets open-source SOTA at its scale on complex chart analysis.The report identifies chart analysis as a particular strength for visually dense content.Source: §1 · p. 2

2 Architecture

• Ovis2.5 retains the foundational Ovis architecture (Lu et al., 2024b), comprising a Visual Tokenizer, Visual Embedding Table, and LLM.Figure 2 illustrates the overall architecture.Source: §2 · p. 3, Figure 2 · p. 3
• The Visual Tokenizer extracts image-patch features and projects them onto a discrete visual-word vocabulary, producing probability distributions over visual tokens.It uses a transformer-based component and a visual head.Source: §2 · p. 3
• The Visual Embedding Table stores embeddings for visual words, and each visual-token embedding is their probability-weighted expected value.This design alleviates structural mismatches between visual and textual modalities.Source: §2 · p. 3
• The pretrained open-source LLM performs cross-modal understanding from visual and textual embeddings to generate text.Source: §2 · p. 3
• Ovis2.5 replaces fixed-resolution image splitting with a native-resolution ViT (NaViT), preserving global structure and fine detail, especially in high-resolution images such as complex charts.RoPE is integrated into every ViT block to reinforce spatial awareness; NaViT is initialized from siglip2-so400m-patch16-5121.Source: §2 · p. 4
• Ovis2.5 replaces its Qwen2.5 backbone with Qwen3 to boost complex-task performance and overall multimodal proficiency.The report attributes the upgrade to Qwen3’s deep-reasoning capabilities.Source: §2 · p. 4

3 Model Training

• The pipeline comprises three pre-training phases followed by two post-training phases, each with distinct configurations and data compositions.Table 1 provides a high-level overview of the training process.Source: §3 · p. 4

3.1.1 Data Composition

• A curated public and in-house multimodal corpus supports training, with OCR samples filtered for resolution, language, and scene diversity and grounding data scaled using detector-localized entities.For reasoning, the dataset combines vanilla CoT with thinking-style data teaching reflection and self-correction, and uses cross-verification by multiple MLLMs to select high-quality synthesized samples.Source: §3.1.1 · pp. 4–5

3.1.2 Training Strategy

• Pre-training progresses from VET initialization to full multimodal training and instruction tuning, expanding resolution and task coverage while adding reflection and self-correction.P1 trains the final ViT layer, visual head, and VET with dynamic position interpolation; P2 trains all modules with RoPE across resolutions up to 17922 total pixels; P3 adds text-only, multi-image, and video inputs alongside thinking-style samples.Source: §3.1.2 · p. 5

3.2.1 Data Composition

• Preference data spans text, image, and video modalities, while RLVR emphasizes mathematics and also includes science and visual reasoning tasks.Reasoning preferences are scored against ground-truth answers; general-purpose preferences use MLLM scoring. RLVR data undergoes quality and offline-difficulty filtering, and some multiple-choice questions are converted to fill-in-the-blank to discourage guessing.Source: §3.2.1 · p. 5

3.2.2 Training Strategy

• Post-training first applies full-parameter multimodal DPO, then uses GRPO to improve reasoning while freezing vision modules.DPO is augmented with an auxiliary NLL objective; GRPO updates only LLM parameters on verifiable, reasoning-centric tasks to preserve general multimodal abilities.Source: §3.2.2 · pp. 5–6

3.3 Infrastructure

• Targeted data-processing and parallelism optimizations improve Ovis2.5 end-to-end training speed by 3 to 4 times, addressing multimodal data-size load imbalance and prior-framework memory limits.Data packing minimizes padding and balances GPU workloads, directly improving training throughput; the Megatron-based framework combines data, tensor, and context parallelism to reduce memory footprint.Source: §3.3 · p. 6

4.1 Overall Performance

• Ovis2.5-9B averages 78.3 and Ovis2.5-2B averages 73.9 on OpenCompass, achieving state-of-the-art performance among open-source models at their respective scales.Ovis2.5-9B scores 71.2 on MMMU, 87.9 on OCRBench, and 65.1 on HallusionBench; the results surpass Keye-VL and GLM-4.1V, narrow the gap to Gemini-2.5, and improve over Ovis2.Source: §4.1 · pp. 6–7, Table 3 · p. 7

4.2 Multimodal Reasoning

• Ovis2.5-9B ranks first on MathVista and WeMath, places among the top two on other math benchmarks, and scores 71.2 on MMMU and 54.4 on MMMU-Pro.Ovis2.5-2B achieves state-of-the-art results in its size category, with substantial gains over comparable open-source models on general reasoning and math-oriented tasks.Source: §4.2 · p. 7, Table 4 · p. 7, Table 5 · p. 8

4.3 OCR & Chart

• Ovis2.5 surpasses leading open-source competitors and GPT-4o on OCRBench v2, and Ovis2.5-9B leads open-source peers on average across chart and document benchmarks.The chart and document evaluation includes ChartQA Pro, ChartQA, DocVQA, and TextVQA.Source: §4.3 · pp. 7–8, Table 6 · p. 8

4.4 Grounding

• Ovis2.5 achieves a state-of-the-art average score of 90.1 among compared open-source models on visual grounding, outperforming all competitors on both RefCOCOg validation and test sets.It also ranks first or second across nearly all RefCOCO and RefCOCO+ splits.Source: §4.4 · p. 8, Table 7 · p. 9

4.5 Multi-Image and Video

• Ovis2.5 scores 67.3 on BLINK and 69.3 on MMT-Bench, outperforming other prominent models, and maintains strong, consistent performance across video benchmarks.The video evaluation spans VideoMME, MVBench, MLVU, and TempCompass, testing sequence-level spatial and temporal understanding.Source: §4.5 · pp. 8–9, Table 8 · p. 9

• Ovis2.5 establishes a new state-of-the-art on OpenCompass for its parameter class, with leading results in STEM, OCR and chart, grounding, and multi-image/video comprehension.Source: §5 · p. 9
• Its NaViT processes images at native, variable resolutions, preserving fine details and global structure important for visually dense content such as charts.Source: §5 · p. 9
• Reflection-oriented training enables thinking mode to self-correct and revise on complex tasks, trading inference latency for accuracy.Source: §5 · p. 9
• Future work targets accurate 4K-level perception, richer temporal reasoning for long-input video, and tighter tool integration for action-augmented reasoning.Source: §5 · p. 9

• The section uses responses to diverse prompts as qualitative examples of Ovis2.5’s capabilities and strengths.Source: A Qualitative Examples · p. 14

A.1 World Knowledge

• The model identifies the plant as a baobab tree (Adansonia digitata) and describes its distinctive features and ecological and cultural roles.Source: §A.1 · p. 14
• The model reasons that the coastal image depicts Istanbul near the Bosphorus, citing ferry-like boats, hilly built-up terrain, and busy waterways.Source: §A.1 · p. 15
• Although other coastal regions could share these traits, the model judges Istanbul the most plausible location based on the combination of water traffic, topography, and architecture.Source: §A.1 · p. 15

A.2 Math Problem Solving

• Belinda’s prompt asks how many toothpicks are added from the 30th pattern to the 31st; the response begins analyzing the count pattern and gives counts of 4 and 12 for n=1 and n=2.The excerpt marks the remaining reasoning as omitted.Source: §A.2 · p. 16
• For an n × n square grid, the response counts 2n(n+1) toothpicks by adding the horizontal and vertical segments.It checks the formula against the first two patterns, which contain 4 and 12 toothpicks, respectively.Source: §A.2 · p. 16

Step 1: Identify the formula for the n-th pattern

• In an n × n grid, each of the n + 1 lines in either orientation has n segments, so each orientation contributes n(n + 1) toothpicks.Source: Step 1: Identify the formula for the… · p. 16
• The n-th pattern contains 2n(n + 1) toothpicks, adding n(n + 1) horizontal and n(n + 1) vertical toothpicks.Each orientation has n + 1 lines with n segments per line.Source: Step 1: Identify the formula for the… · p. 16

Step 2: Find the difference between consecutive patterns

• Subtracting T(n) = 2n(n + 1) from T(n + 1) = 2(n + 1)(n + 2) simplifies to 4(n + 1) added toothpicks.Source: Step 2: Find the difference between… · p. 16

• The transition calculation substitutes n = 30 into the difference formula.Source: Step 3: Apply the formula to the 30th… · p. 16
• 124 toothpicks are added to the 30th pattern to make the 31st.Source: Step 3: Apply the formula to the 30th… · p. 16

Step 5: Find the Radius of the Big Circle

• The geometry solution derives a small-circle radius of 6 and a big-circle radius of 18 from the 1:3 radius ratio and tangent-chord setup.It uses the diameter to establish a right angle in △ABC, then applies Pythagoras.Source: Step 5: Find the Radius of the Big… · p. 17, Step 4: Apply Pythagoras to △ABC In… · p. 17, Step 1: Define Variables and Key… · p. 17

Final Directions

• The physics example gives a reflected-ray angle of 60° and a refracted-ray angle of approximately 49.3° with the normal, bending toward the normal in glass.The reflection result places the ray on the same side as the incident ray.Source: Final Directions · p. 18, Step 1: Reflected Ray (Law of… · p. 18

Step 3: Determine Genotype of Individual 9

• For the pedigree problem, the model identifies Individual 9 as a carrier with genotype XAXa, given her unaffected status and parents’ genotypes.The example treats the inheritance pattern as X-linked recessive.Source: Step 3: Determine Genotype of… · p. 19, Step 1: Determine Mode of Inheritance · p. 19

A.5 OCR

• The OCR examples return requested document fields as key-value pairs and provide word-level text localization with bounding boxes.The extracted fields include the lot number, study director, container count, handling requirements, physical description, gross weight, shipping method, and date; the localization example boxes words including PIZZA and BREWING.Source: §A.5 · pp. 21–22

A.7 Grounding

• Grounding examples provide a bounding box for a green tent, boxes for described image objects, and point coordinates for hot-air balloons.The grounded caption names three birds, the moon, and tree branches with leaves.Source: §A.7 · pp. 26–27

A.8 Multi-Images

• In a missing-piece comparison, the model selects the third image as the missing lower-right portion of the first image.Source: §A.8 · p. 28

A.9 Video

• The video example’s response begins by describing a live musical performance featuring two musicians on stage.Source: §A.9 · p. 30
• The video description centers on a saxophonist performing before a red curtain, with a guitarist accompanying him and focused lighting shaping an intimate live-stage scene.The saxophonist remains the central focus, while the guitarist plays a secondary but complementary role.Source: §A.9 · p. 30, Figure 22