Read a video, then read Figure 2 from a PDF Claude Code

{
  "video_path": "/path/to/Qwen-3.8-VL_v4.mp4"
}

Video: /path/to/Qwen-3.8-VL_v4.mp4 | 150.2s | 300 frames @ 2.0fps [0:00.0–2:30.2] | 384x672 (HxW)

{
  "file_path": "/path/to/qwen3vl.pdf",
  "pages": "1-6"
}

pdftoppm is not installed. Install poppler-utils (e.g. `brew install poppler` or `apt-get install poppler-utils`) to enable PDF page rendering.

{
  "file_path": "/path/to/qwen3vl.pdf",
  "pages": "1-6",
  "budget": "large"
}

[PDF Start]
Total pages: 35 | Showing pages: 1-6
[Page 1 View] 1728x1216 (HxW)

[Page 1 Extracted Text]
July 25, 2026
Qwen3-VL Technical Report
Qwen Team
https://chat.qwen.ai
https://huggingface.co/Qwen
https://modelscope.cn/organization/qwen
https://github.com/QwenLM/Qwen3-VL
Abstract
We introduce Qwen3-VL, the most capable vision–language model in the Qwen series to
date, achieving superior performance across a broad range of multimodal benchmarks.
It natively supports interleaved contexts of up to 256K tokens, seamlessly integrating text, images, and video. The model family includes both dense (2B/4B/8B/32B)
and mixture-of-experts (30B-A3B/235B-A22B) variants to accommodate diverse latency–quality trade-offs. Qwen3-VL delivers three core pillars: (i) markedly stronger
pure-text understanding, surpassing comparable text-only backbones in several cases;
(ii) robust long-context comprehension with a native 256K-token window for both text
and interleaved multimodal inputs, enabling faithful retention, retrieval, and crossreferencing across long documents and videos; and (iii) advanced multimodal reasoning
across single-image, multi-image, and video tasks, demonstrating leading performance
on comprehensive evaluations such as MMMU and visual-math benchmarks (e.g., MathVista and MathVision). Architecturally, we introduce three key upgrades: (i) an enhanced
interleaved-MRoPE for stronger spatial–temporal modeling across images and video; (ii)
DeepStack integration, which effectively leverages multi-level ViT features to tighten
vision–language alignment; and (iii) text-based time alignment for video, evolving from
T-RoPE to explicit textual timestamp alignment for more precise temporal grounding. To
balance text-only and multimodal learning objectives, we apply square-root reweighting, which boosts multimodal performance without compromising text capabilities.
We extend pretraining to a context length of 256K tokens and bifurcate post-training
into non-thinking and thinking variants to address distinct application requirements.
Furthermore, we allocate additional compute resources to the post-training phase to
further enhance model performance. Under comparable token budgets and latency
constraints, Qwen3-VL achieves superior performance in both dense and Mixture-ofExperts (MoE) architectures. We envision Qwen3-VL serving as a foundational engine for
image-grounded reasoning, agentic decision-making, and multimodal code intelligence
in real-world workflows.
1
[Page 2 View] 1728x1216 (HxW)

[Page 2 Extracted Text]
1 Introduction
Vision–language models (VLMs) have achieved substantive progress in recent years, evolving from
foundational visual perception to advanced multimodal reasoning across images and video. The rapid
advancement of VLMs has given rise to a rapidly expanding landscape of downstream applications—such
as long-context understanding, STEM reasoning, GUI comprehension and interaction, and agentic
workflows. Crucially, these advances must not erode the underlying large language model’s (LLM’s)
linguistic proficiency; multimodal models are expected to match or surpass their text-only counterparts
on language benchmarks.
In this report, we present Qwen3-VL and its advances in both general-purpose and advanced applications.
Built on the Qwen3 series (?), we instantiate four dense models (2B/4B/8B/32B) and two mixture-ofexperts (MoE) models (30B-A3B / 235B-A22B), each trained with a context window of up to 256K tokens
to enable long-context understanding. By optimizing the training corpus and training strategy, we
preserve the underlying LLM’s language proficiency during vision–language (VL) training, thereby
substantially improving overall capability. We release both non-thinking and thinking variants; the latter
demonstrates significantly stronger multimodal reasoning capabilities, achieving superior performance
on complex reasoning tasks.
We first introduce the architectural improvements, which span three components: 1) Enhanced positional
encoding. In Qwen2.5-VL, we used MRoPE as a unified positional encoding scheme for text and vision.
We observed that chunking the embedding dimensions into temporal (t), horizontal (h), and vertical (w)
groups induces an imbalanced frequency spectrum and hampers long-video understanding. We therefore
adopt an interleaved MRoPE that distributes t, h, and w uniformly across low- and high-frequency bands,
yielding more faithful positional representations. 2) DeepStack for cross-layer fusion. To strengthen
vision–language alignment, we incorporate the pioneering DeepStack (?) mechanism. Visual tokens
from different layers of the vision encoder are routed to corresponding LLM layers via lightweight
residual connections, enhancing multi-level fusion without introducing extra context length. 3) Explicit
video timestamps. We replace the absolute-time alignment via positional encoding used in Qwen2.5-VL
with explicit timestamp tokens to mark frame groups, providing a simpler and more direct temporal
representation. In addition, on the optimization side, we move from a per-sample loss to a square-rootnormalized per-token loss, which better balances the contributions of text and multimodal data during
training.
To build a more capable and robust vision–language foundation model, we overhauled our training
data in terms of quality, diversity, and structure. Key upgrades include enhanced caption supervision,
expanded omni-recognition and OCR coverage, normalized grounding with 3D/spatial reasoning, and
new corpora for code, long documents, and temporally grounded video. We further infused chain-ofthought reasoning and high-quality, diverse GUI-agent interaction data to bridge perception, reasoning,
and action. Together, these innovations enable stronger multimodal understanding, precise grounding,
and tool-augmented intelligence.
Our training pipeline consists of two stages: pretraining and post-training. Pretraining proceeds in
four phases: a warm-up alignment phase that updates only the merger (vision–language projection)
layers while keeping the rest of the model frozen, followed by full-parameter training with progressively
larger context windows at 8K, 32K, and 256K sequence lengths. Post-training comprises three phases: (i)
supervised fine-tuning on long chain-of-thought data, (ii) knowledge distillation from stronger teacher
models, and (iii) reinforcement learning.
The above innovations equip Qwen3-VL with strong capabilities not only as a robust vision–language
foundation model but also as a flexible platform for real-world multimodal intelligence—seamlessly integrating perception, reasoning, and action across diverse application domains. In the following sections,
we present the model architecture, training framework, and extensive evaluations that demonstrate its
consistent and competitive performance on text, vision, and multimodal reasoning benchmarks.
2 Model Architecture
Following Qwen2.5-VL (?), Qwen3-VL adopts a three-module architecture comprising a vision encoder,
an MLP-based vision–language merger, and a large language model (LLM). Figure 1 depicts the detailed
model structure.
Large Language Model: Qwen3-VL is instantiated in three dense variants (Qwen3-VL-2B/4B/8B/32B)
and two MoE variants (Qwen3-VL-30B-A3B, Qwen3-VL-235B-A22B), all built upon Qwen3 backbones.
The flagship model, Qwen3-VL-235B-A22B, has 235B total parameters with 22B activated per token. It
2
[Page 3 View] 1728x1216 (HxW)

[Page 3 Extracted Text]
Figure 1: The Qwen3-VL framework integrates a vision encoder and a language model decoder to process
multimodal inputs, including text, images, and video. The vision encoder is specifically designed to
handle dynamic, native-resolution visual inputs, mapping them to visual tokens of variable length.
To enhance perceptual capability and preserve rich visual information, we incorporate the pioneering
DeepStack mechanism, which injects visual tokens from multiple layers of the vision encoder into
corresponding layers of the LLM. Furthermore, we adopt Interleaved MRoPE to encode positional
information for multimodal inputs with a balanced frequency spectrum, and introduce text-based
timestamp tokens to more effectively capture the temporal structure of video sequences.
outperforms most VLMs across a broad set of multimodal tasks and surpasses its text-only counterpart
on the majority of language benchmarks.
Vision Encoder: We utilize the SigLIP-2 architecture (?) as our vision encoder and continue training
it with dynamic input resolutions, initialized from official pretrained checkpoints. To accommodate
dynamic resolutions effectively, we employ 2D-RoPE and interpolate absolute position embeddings based
on input size, following the methodology of CoMP (?). Specifically, we default to the SigLIP2-SO-400M
variant and use SigLIP2-Large (300M) for small-scale LLMs (2B and 4B).
MLP-based Vision-Language Merger: As in Qwen2.5-VL, we use a two-layer MLP to compress 2 × 2
visual features from the vision encoder into a single visual token, aligned with the LLM’s hidden
dimension. Additionally, we deploy specialized mergers to support the DeepStack mechanism (?), the
details of which are fully described in Section 2.2.
2.1 Interleaved MRoPE
Qwen2-VL (?) introduced MRoPE to model positional information for multimodal inputs. In its original
formulation, the embedding dimensions are partitioned into temporal (t), horizontal (h), and vertical (w)
subspaces, each assigned distinct rotary frequencies. This results in an imbalanced frequency spectrum,
which subsequent studies have shown to degrade performance on long-video understanding benchmarks.
To address this, we redesign the frequency allocation by interleaving the t, h, and w components across
the embedding dimensions (?). This ensures that each spatial–temporal axis is uniformly represented
across both low- and high-frequency bands. The resulting balanced spectrum mitigates the original
spectral bias and significantly improves long-range positional modeling for video.
2.2 DeepStack
We draw inspiration from DeepStack (?) and inject visual tokens into multiple layers of the LLM.
Unlike the original DeepStack approach, which stacks tokens from multi-scale visual inputs, we extend
3
[Page 4 View] 1728x1216 (HxW)

[Page 4 Extracted Text]
DeepStack to extract visual tokens from intermediate layers of the Vision Transformer (ViT). This design
preserves rich visual information, ranging from low- to high-level representations.
Specifically, as illustrated in Figure 1, we select features from three distinct levels of the vision encoder.
Subsequently, dedicated vision–language merger modules project these multi-level features into visual
tokens, which are then added directly to the corresponding hidden states of the first three LLM layers.
2.3 Video Timestamp
In Qwen2.5-VL, a time-synchronized variant of MRoPE is employed to endow the model with temporal
awareness. However, we identify two key limitations of this approach: (1) By tying temporal position
IDs directly to absolute time, the method produces excessively large and sparse temporal position ids for
long videos, degrading the model’s ability to understand long temporal contexts. (2) Effective learning
under this scheme requires extensive and uniformly distributed sampling across various frame rates
(fps), significantly increasing the cost of training data construction.
To address these issues, we adopt a textual token–based time encoding strategy (?), wherein each video
temporal patch is prefixed with a timestamp expressed as a formatted text string—e.g., <3.0 seconds>.
Furthermore, during training, we generate timestamps in both seconds and HMS (hours:minutes:seconds)
formats to ensure the model learns to interpret diverse timecode representations. Although this approach
incurs a modest increase in context length, it enables the model to perceive temporal information more
effectively and precisely, thereby facilitating time-aware video tasks such as video grounding and dense
captioning.
3 Pre-Training
3.1 Training Recipe
We first enhance the vision encoder by conducting continuous training with dynamic resolutions based
on the pre-trained SigLIP-2 model. The overall Qwen3-VL model adopts a three-module architecture,
comprising this vision encoder, an MLP-based vision–language merger, and a Qwen3 large language
model (LLM) backbone. Building on this architecture, our pre-training methodology is systematically
structured into four distinct stages, designed to progressively build capabilities from basic alignment to
long-context understanding. An overview of these stages is presented in Table 1.
Table 1: Training setup and hyperparameters across different stages for Qwen3-VL.
Stage Objective Training Token Budget Sequence Length
S0 Vision-Language Alignment Merger 67B 8,192
S1 Multimodal Pre-Training All ~1T 8,192
S2 Long-Context Pre-Training All ~1T 32,768
S3 Ultra-Long-Context Adaptation All 100B 262,144
Stage 0: Vision-Language Alignment. The initial stage (S0) focuses on efficiently bridging the modality
gap between the vision encoder and the LLM. Crucially, only the parameters of the MLP merger are
trained during this phase, while both the vision encoder and the LLM backbone remain frozen. We utilize
a curated dataset of approximately 67B tokens, consisting of high-quality image-caption pairs, visual
knowledge collections, and optical character recognition (OCR) data. All training is conducted with a
sequence length of 8,192. This alignment-first approach establishes a solid foundation for cross-modal
understanding before proceeding to full-parameter training.
Stage 1: Multimodal Pre-Training. Following the initial alignment, Stage 1 (S1) transitions to fullparameter Multimodal Pre-Training. In this phase, we unfreeze all model components—the vision
encoder, the merger, and the LLM—for joint end-to-end training. The model is trained on a massive
and diverse dataset of approximately 1 trillion (1T) tokens. To maintain the LLM’s strong language
abilities, the data mixture is composed of vision-language (VL) data and text-only data. The VL portion
is rich and varied, adding interleaved image-text documents, visual grounding tasks, visual question
answering (VQA), data from STEM domains, and a small amount of video data to introduce temporal
understanding. The sequence length remains at 8,192.
Stage 2: Long-Context Pre-Training. Stage 2 (S2) aims to significantly extend the model’s contextual
processing abilities. A key change in this stage is the quadrupling of the sequence length to 32,768, while
4
[Page 5 View] 1728x1216 (HxW)

[Page 5 Extracted Text]
all model parameters continue to be trainable. Training is conducted on a dataset of approximately
1T tokens, with an adjusted data mixture to support long-context tasks. The proportion of text-only
data is increased to bolster long-form text comprehension, while the remaining VL data incorporates a
significantly larger volume of video and agent-oriented instruction-following data. This stage is critical
for enabling the model to process and reason over longer videos and complex, multi-step tasks.
Stage 3: Ultra-Long-Context Adaptation. The final stage (S3) is a specialized phase designed to push the
model’s context window to its operational limits. Here, we dramatically increase the sequence length to
262,144. The model is trained on a more focused 100B token dataset specifically curated for this purpose.
The data is also composed of text-only data and VL data, with a strong emphasis on long-video and longdocument understanding tasks. This final adaptation solidifies Qwen3-VL’s proficiency in processing
and analyzing extremely long sequential inputs, a key capability for applications like comprehensive
document analysis and lengthy video summarization.
3.2 Pre-Training Data
3.2.1 Image Caption and Interleaved Text-Image Data
To build a robust foundation model for general-purpose vision–language understanding, we significantly
expand and refine two core data modalities: image–caption pairs and interleaved text–image sequences.
Our strategy emphasizes high-quality, diverse, and semantically rich multimodal grounding, supported
by purpose-built models and rigorous filtering pipelines.
Image Caption Data: We curate a large-scale corpus of contemporary, predominantly Chinese–English
multilingual image–text pairs from web sources and apply a multi-stage refinement pipeline centered
on a specialized Qwen2.5-VL-32B model fine-tuned for recaptioning. This model leverages the original
raw text associated with each image to generate more comprehensive, fluent, and fine-grained captions—enriching descriptions of visual elements (e.g., object attributes, spatial layouts, and contextual
semantics) while simultaneously improving the linguistic quality and informativeness of the textual
component.
Deduplication is performed exclusively on the recaptioned text using semantic similarity metrics, ensuring removal of redundant samples without sacrificing visual diversity. To further enhance coverage of
underrepresented concepts, we apply clustering (???) over visual embeddings to identify sparse regions
in the data distribution and perform targeted augmentation. The result is a high-fidelity caption dataset
that balances scale, diversity, and descriptive granularity.
Interleaved Text-Image Data: We collect diverse real-world multimodal documents sourced from recent
Chinese and English websites (???). All documents undergo domain classification (?) using a lightweight
Qwen-based scorer fine-tuned for fine-grained domain identification. Based on validation experiments
across domains, we systematically exclude harmful or low-value categories—such as advertisements,
promotional content, and clickbait—using the same efficient scorer to filter out undesirable samples.
For book-scale interleaved data, we employ a fine-tuned Qwen2.5-VL-7B model to perform high-accuracy
multimodal parsing, precisely extracting and aligning text with embedded figures, diagrams, and
photographs. To enable ultra-long context modeling, we construct a specialized subset by merging
consecutive pages into sequences of up to 256K tokens, preserving natural page order and multimodal
coherence. During preprocessing, we enforce strict quality controls: (i) pure-text or low-alignment
segments are removed; (ii) for ultra-long book sequences, we require a minimum page count and a
minimum image-to-text ratio to ensure meaningful visual–textual interaction throughout the context. This
yields a clean, diverse, and layout-aware interleaved corpus optimized for both grounded understanding
and long-range multimodal reasoning.
3.2.2 Knowledge
World knowledge is essential for multimodal large language models (MLLMs) to achieve robust visual
understanding, grounded reasoning, and entity-aware generation across diverse downstream tasks. To
equip Qwen3-VL with a comprehensive grasp of both real-world and fictional concepts, we construct a
large-scale pretraining dataset centered on well-defined entities spanning more than a dozen semantic
categories—including animals, plants, landmarks, food, and everyday objects such as vehicles, electronics,
and clothing.
Real-world entities follow a long-tailed distribution: prominent concepts appear frequently with highquality annotations, while the majority are rare. To address this imbalance, we adopt an importance-based
sampling strategy. High-prominence entities are sampled more heavily to ensure a sufficient learning
5
[Page 6 View] 1728x1216 (HxW)

[Page 6 Extracted Text]
signal, while low-prominence entities are included in smaller proportions to maintain broad coverage
without overwhelming the training process. This approach effectively balances data quality, utility, and
diversity.
All retained samples undergo a multi-stage refinement pipeline. In addition to standard filtering for
noise and misalignment, we replace original or sparse captions—such as generic alt-text—with richer,
LLM-generated descriptions. These enhanced captions not only identify the main entity but also describe
its visual attributes, surrounding context, spatial layout, and interactions with other objects or people,
thereby providing a more complete and grounded textual representation.
Together, these efforts yield a knowledge-rich, context-aware, and discrimination-focused training signal
that significantly enhances Qwen3-VL’s ability to recognize, reason about, and accurately describe visual
concepts in real-world scenarios.
3.2.3 OCR, Document Parsing and Long Document Understanding
OCR: To enhance OCR performance on real-world images, we curate a dataset of 30 million in-house
collected samples using a coarse-to-fine pipeline. This pipeline refines OCR annotations by integrating
pseudo-labels from OCR-specialized models with refinements from Qwen2.5-VL—without any human
annotation. Expanding beyond the 10 languages supported by Qwen2.5-VL (excluding Chinese and
English), we incorporate an additional 29 languages, synthesizing approximately 30 million high-quality
multilingual OCR samples and curating over 1 million internal real-world multilingual images.
Document Parsing: For document parsing, we collect 3 million PDFs from Common Crawl, evenly
distributed across 10 document types (300K samples each), along with 4 million internal documents. An
in-house layout model first predicts the reading order and bounding boxes for textual and non-textual
regions; Qwen2.5-VL-72B then performs region-specific recognition. The outputs are reassembled into
position-aware, layout-aligned parsing data.
To ensure robust parsing across heterogeneous formats, we design a unified annotation framework
supporting two representations:
• QwenVL-HTML, which includes fine-grained, element-level bounding boxes;
• QwenVL-Markdown, where only images and tables are localized, with tables encoded in LaTeX.
We construct a large-scale synthetic HTML corpus with precise annotations and systematically convert it
to Markdown format. To further improve model generalization, we generate pseudo-labels on extensive
collections of real documents and filter them for quality. The final training set combines synthetic and
high-quality pseudo-labeled data to enhance both scalability and robustness.
Long Document Understanding: To enhance the model’s ability to understand multi-page PDFs—often
spanning dozens of pages—we leverage a large-scale corpus of long-document data. First, we synthesize
long-document parsing sequences by merging single-page document samples. In each sequence, multiple
page images are placed at the beginning, followed by their corresponding text derived from OCR or
HTML parsing. Second, we construct long-document visual question answering (VQA) data. Specifically,
we sample high-quality multi-page PDFs and generate a diverse set of VQA examples that require the
model to reason across multiple pages and heterogeneous document elements—such as charts, tables,
figures, and body text. We carefully balance the distribution of question types and ensure that supporting
evidence draws from a wide range of modalities and layout components, thereby promoting robust,
grounded, and multi-hop reasoning over extended contexts.
3.2.4 Grounding and Counting
Visual grounding is a fundamental capability for multimodal models, enabling them to accurately identify,
interpret, and localize a wide spectrum of visual targets from specific objects to arbitrary image regions.
In Qwen3-VL, we systematically enhance grounding proficiency and support two grounding modalities:
bounding boxes and points. These representations allow for precise and flexible interpretation of image
content across diverse scenarios and downstream tasks. In addition, we extend the grounding capacity of
the model to support counting, enabling quantitative reasoning about visual entities. In the following,
we briefly describe the data construction pipelines for grounding and counting.
Box-based Grounding: We begin by aggregating widely used open-source datasets, including COCO (?),
Objects365 (?), OpenImages (?), and RefCOCO/+/g (??). To further enrich data diversity, we developed
an automated synthesis pipeline that generates high-quality object annotations across a broad range
of scenarios. This pipeline operates in three stages: (i) object candidates are extracted from unlabeled
6
[PDF End]

{
  "file_path": "/path/to/qwen3vl.pdf",
  "pages": "7-12",
  "budget": "large"
}

[PDF Start]
Total pages: 35 | Showing pages: 7-12
[Page 7 View] 1728x1216 (HxW)

[Page 7 Extracted Text]
images using Qwen2.5-VL; (ii) these candidates are localized and annotated using both open-vocabulary
detectors (specifically, Grounding DINO (?)) and Qwen2.5-VL; and (iii) the resulting annotations undergo
quality assessment, with low-confidence or inaccurate ones systematically filtered out. Through this
approach, we constructed a large-scale, highly diverse box-based grounding dataset spanning a wide
variety of visual contexts and object categories.
Point-based Grounding: To ensure robust point-based grounding, we curated a comprehensive dataset
combining publicly available and synthetically generated pointing annotations. It integrates three
sources: (i) public pointing and counting annotations from PixMo (?); (ii) object grounding data derived
from public object detection and instance segmentation benchmarks; and (iii) high-precision pointing
annotations generated by a dedicated synthesis pipeline designed to target fine-grained image details.
Counting: Building upon the grounding data, we curated a high-quality subset to form the basis of our
counting dataset, which includes three distinct task formulations: direct counting, box-based counting,
and point-based counting. Collectively, these three task types constitute a comprehensive counting
dataset.
Different from Qwen2.5-VL, we adopt a normalized coordinate system scaled to the range [0, 1000] in this
version. This design improves robustness to variations in image resolution and aspect ratio across diverse
inputs, while also simplifying post-processing and enhancing the usability of predicted coordinates in
downstream applications.
3.2.5 Spatial Understanding and 3D Recognition
To facilitate sophisticated interaction with the physical world, Qwen3-VL is designed with a deep
understanding of spatial context. This enables the model to interpret spatial relationships, infer object
affordances, and perform action planning and embodied reasoning. It can also estimate the 3D spatial
positions of objects from a single monocular image. To support these capabilities, we created two
comprehensive datasets focused on Spatial Understanding and 3D Grounding.
Spatial Understanding. Beyond localizing objects, Qwen3-VL is trained to reason about spatial relationships, object affordances, and feasible actions in 2D scenes—capabilities essential for embodied AI
and interactive applications. To this end, we construct a specialized dataset that goes beyond standard
grounding by incorporating: (i) relational annotations (e.g., “the cup to the left of the laptop”), (ii)
affordance labels (e.g., “graspable”, “pressable”, “sittable”), and (iii) action-conditioned queries that
require planning (e.g., “What should I move first to reach the book behind the monitor?”). These samples
are derived from both curated real-world scenes and synthetically generated layouts, with natural language queries automatically generated via templated and LLM-based methods to ensure diversity and
complexity. Critically, all spatial references are expressed relative to other objects or scene frames, rather
than absolute coordinates, encouraging robust relational reasoning. This training enables Qwen3-VL to
not only answer “where” questions but also “how” and “what can be done” — forming a foundation for
agentic interaction with visual environments.
3D Grounding. To further enhance the model’s ability to understand the physical world from images,
we constructed a specialized pretraining dataset for 3D visual grounding. We sourced data from public
collections of diverse indoor and outdoor scenes and reformulated it into a visual question-answering
format. Each sample consists of: 1) a single-view camera image, 2) a natural language referring expression,
and 3) the corresponding 9-DoF 3D bounding box annotations in a structured JSON format, specifying
the object’s spatial position and semantic label. As the 3D bounding boxes are derived from multiple
sensors and data sources, they exhibit varying camera intrinsic parameters and inherent noise. To this
end, we filter out heavily occluded and inaccurate labels and follow Omni3D (?) to unify all data into a
virtual camera coordinate system. We also synthesized a large corpus of descriptive captions to create
rich textual queries for 3D grounding. These descriptions go beyond naming the object’s category to
include detailed attributes, layout arrangements, spatial location, visual affordances, and interactions
with surrounding objects—yielding more fine-grained and grounded referring expressions.
3.2.6 Code
We enhance the Qwen3-VL series with dedicated coding capabilities by incorporating two categories of
code-related data into the training corpus, enabling the model to read, write, and reason about programs
in both text-only and visually grounded contexts.
Text-Only Coding. We reuse the extensive code corpus from the Qwen3 and Qwen3-Coder series. This
large-scale dataset spans a wide range of programming languages and domains—including software
development, algorithmic problem solving, mathematical reasoning, and agent-oriented tasks—and
7
[Page 8 View] 1728x1216 (HxW)

[Page 8 Extracted Text]
establishes the model’s foundational understanding of code syntax, algorithmic logic, and generalpurpose program generation.
Multimodal Coding. To address tasks requiring both visual understanding and code generation, we
curate data for a diverse suite of multimodal coding tasks. This dataset, sourced from both open-source
datasets and internal synthesis pipelines, teaches the model to jointly understand visual inputs and
generate functional code. The data covers several key tasks, including: converting UI screenshots into
responsive HTML/CSS; generating editable SVG codes from images (?); solving visual programming
challenges (?); answering multimodal coding questions (e.g., StackOverflow posts with images); and transcribing visual representations (such as flowcharts, diagrams, and LATEX equations) into their respective
code or markup. This novel data mixture enables Qwen3-VL to act as a bridge between visual perception
and executable logic.
3.2.7 Video
The video comprehension capabilities of Qwen3-VL have been substantially advanced, enabling robust
modeling of temporal dynamics across frames, fine-grained perception of spatial relationships, and
coherent summarization of ultra-long video sequences. This enhancement is underpinned by a data
processing pipeline featuring two principal innovations:
Temporal-Aware Video Understanding. (i) Dense Caption Synthesis: For long video sequences, we
employ a short-to-long caption synthesis strategy to generate holistic, timestamp-interleaved, and
temporally coherent story-level descriptions. Leveraging in-house captioning models, we further produce
fine-grained annotations that jointly capture event-level temporal summaries and segment-specific visual
details. (ii) Spatio-Temporal Video Grounding: We curate and synthesize large-scale video data annotated
at the levels of objects, actions, and persons to strengthen the model’s spatio-temporal grounding
capabilities, thereby improving its capacity for fine-grained video understanding.
Video Data Balancing and Sampling. (i) Source Balancing: To ensure data balance and diversity, we
assemble a large-scale dataset encompassing various video sources, including instructional content,
cinematic films, egocentric recordings, etc. Dataset balance is achieved through systematic curation
guided by metadata such as video titles, duration, and categorical labels. (ii) Length-Adaptive Sampling:
During pre-training stages, we dynamically adjust sampling parameters, such as frames per second (fps)
and the maximum number of frames, according to different sequence length constraints. This adaptive
strategy mitigates information loss associated with suboptimal sampling practices (e.g., overly sparse
frame selection or excessively low spatial resolution), thus preserving visual details and optimizing
training efficacy.
3.2.8 Science, Technology, Engineering, and Mathematics (STEM)
Multimodal reasoning lies at the heart of Qwen3-VL, with STEM reasoning constituting its most essential
part. Our philosophy follows a divide-and-conquer strategy: we first develop fine-grained visual
perception and robust linguistic reasoning capabilities independently, and then integrate them in a
synergistic manner to achieve effective multimodal reasoning.
Visual Perception Data. We develop a dedicated synthetic data generation pipeline that constructs
geometric diagrams through programmatic (code-based) rendering. Using this pipeline, we generate: (i)
1 million point-grounding samples, such as intersection points, corners, and centers of gravity; and (ii) 2
million perception-oriented visual question answering pairs targeting fine-grained visual understanding
of diagrams. To obtain high-fidelity textual descriptions, we further implement a two-stage captioning
framework: an initial generation phase followed by rigorous model-based verification. Both stages
employ ensembles of specialized models to ensure accuracy and descriptive granularity. This process
yields a comprehensive dataset of 6 million richly annotated diagram captions spanning diverse STEM
disciplines.
Multi-modal Reasoning Data. The majority of our multi-modal reasoning data consists of over 60
million K–12 and undergraduate-level exercises, meticulously curated through a rigorous cleaning and
reformulation pipeline. During quality filtering, we discard low-quality items, including those with
corrupted images, irrelevant content, or incomplete or incorrect answers. During the reformulation stage,
we translate exercises between Chinese and English and standardize the format of answers—such as
step-by-step solution lists, mathematical expressions, and symbolic notations—to ensure consistency
and uniform presentation. Regarding long CoT problem-solving data, we synthesize over 12 million
multimodal reasoning samples paired with images. To ensure the continuity and richness of the reasoning
process, we utilize the original rollouts generated by a strong reasoning model. To guarantee data
reliability and applicability, each sample’s reasoning trajectory undergoes rigorous validation—combining
8
[Page 9 View] 1728x1216 (HxW)

[Page 9 Extracted Text]
rule-based checks and model-based verification—and any instances containing ambiguous answers or
code-switching are explicitly filtered out. Furthermore, to enhance reasoning quality, we retain only
challenging problems via rejection sampling.
Linguistic Reasoning Data. In addition to multimodal reasoning data, we also incorporate reasoning
data from Qwen3, as multimodal reasoning capabilities are largely derived from linguistic reasoning
competence.
3.2.9 Agent
GUI: To endow Qwen3-VL with agentic capability for autonomous interaction with graphical user
interfaces (GUIs), we curate and synthesize large-scale, cross-platform data spanning desktop, mobile, and
web environments (???). For GUI interface perception, we leverage metadata, parsing tools, and human
annotations to construct tasks such as element description, dense captioning, and dense grounding,
enabling robust understanding of diverse user interfaces. For agentic capability, we assemble multistep task trajectories via a self-evolving trajectory-production framework, complemented by targeted
human audits; we also carefully design and augment Chain-of-Thought rationales to strengthen planning,
decision-making, and reflective self-correction during real-world execution.
Function Calling: For general function calling capabilities with multimodal contexts, we build a multimodal function calling trajectory synthesis pipeline. We first instruct capable models with images to
generate user queries and their corresponding function definitions. We then sample model function calls
with rationales and synthesize the function responses. This process is repeated until the user’s query is
judged to be solved. Between each step, trajectories can be filtered out due to formatting errors. Such a
pipeline enables us to construct large-scale multimodal function-calling trajectories from vast images,
without the need to implement executable functions.
Search: Among the general function calling capabilities, we regard the ability to perform searches as key
to facilitating knowledge integration for long-tail entities in real-world scenarios. In this case, we collect
multimodal factual lookup trajectories with online image search and text search tools, encouraging the
model to perform searches for unfamiliar entities. By doing so, the model learns to gather information
from the web to generate more accurate responses.
4 Post-Training
4.1 Training Recipe
Our post-training pipeline is a three-stage process designed to refine the model’s instruction-following
capabilities, bolster its reasoning abilities, and align it with human preferences. The specific data and
methods for each stage are detailed in the subsequent sections.
Supervised Fine-Tuning (SFT). The first stage imparts instruction-following abilities and activates latent
reasoning skills. This is conducted in two phases: an initial phase at a 32k context length, followed by
an extension to a 256k context window that focuses on long-document and long-video data. To cater
to different needs, we bifurcate the training data into standard formats for non-thinking models and
Chain-of-Thought (CoT) formats for thinking models, the latter of which explicitly models the reasoning
process.
Strong-to-Weak Distillation. The second stage employs knowledge distillation, where a powerful
teacher model transfers its capabilities to our student models. Crucially, we perform this distillation using
text-only data to fine-tune the LLM backbone. This method proves highly effective, yielding significant
improvements in reasoning abilities across both text-centric and multimodal tasks.
Reinforcement Learning (RL). The final stage utilizes RL to further enhance model performance and
alignment. This phase is divided into Reasoning RL and General RL. We apply large-scale reinforcement
learning across a comprehensive set of text and multimodal domains, including but not limited to math,
OCR, grounding, and instruction-following, to improve finer-grained capabilities.
4.2 Cold Start Data
4.2.1 SFT Data
Our principal objective is to endow the model with the capacity to address a wide spectrum of realworld scenarios. Building upon the foundational capabilities of Qwen2.5-VL, which is proficient in
9
[Page 10 View] 1728x1216 (HxW)

[Page 10 Extracted Text]
approximately eight core domains and 30 fine-grained subcategories, we have strategically expanded
its functional scope. This expansion was achieved by integrating insights from community feedback,
academic literature, and practical applications, facilitating the introduction of novel capabilities. These
include, but are not limited to, spatial reasoning for embodied intelligence, image-grounded reasoning for
fine-grained visual understanding, spatio-temporal grounding in videos for robust object tracking, and
the comprehension of long-context technical documents spanning hundreds of pages. Guided by these
target tasks and grounded in authentic use cases, we systematically curated the SFT dataset through the
meticulous selection and synthesis of samples from open-source datasets and web resources. This targeted
data engineering effort has been instrumental in establishing Qwen3-VL as a more comprehensive and
robust multimodal foundation model.
This dataset comprises approximately 1,200,000 samples, strategically composed to foster robust multimodal capabilities. This collection is partitioned into unimodal and multimodal data, with one-third
consisting of text-only entries and the remaining two-thirds comprising image-text and video-text pairs.
The integration of multimodal content is specifically designed to enable the model to interpret complex,
real-world scenarios. To ensure global relevance, the dataset extends beyond its primary Chinese and English corpora to include a diverse set of multilingual samples, thereby broadening its linguistic coverage.
Furthermore, it simulates realistic conversational dynamics by incorporating both single-turn and multiturn dialogues contextualized within various visual settings, from single-image to multi-image sequences.
Crucially, the dataset also features interleaved image-text examples engineered to support advanced
agentic behaviors, such as tool-augmented image search and visually-grounded reasoning. This heterogeneous data composition ensures comprehensive coverage and enhances the dataset’s representativeness
for training generalizable and sophisticated multimodal agents.
Given Qwen3-VL’s native support for a 256K token context length, we employ a staged training strategy
to optimize for computational efficiency. This strategy comprises two phases: an initial one-epoch
training phase with a sequence length of 32K tokens, followed by a second epoch at the full 256K token
length. During this latter stage, the model is trained on a curriculum that interleaves long-context inputs
with data sampled at the 32K token length. The long-context inputs include materials such as hundreds
of pages of technical documents, entire textbooks, and videos up to two hours in duration.
The quality of training data is a critical determinant of the performance of vision-language models.
Datasets derived from open-source and synthetic origins are often plagued by substantial variability
and noise, including redundant, irrelevant, or low-quality samples. To mitigate these deficiencies, the
implementation of a rigorous data filtering protocol is indispensable. Accordingly, our data curation
process incorporates a two-phase filtering pipeline: Query Filtering and Response Filtering.
Query Filtering. In this initial phase, we leverage Qwen2.5-VL to identify and discard queries that are
not readily verifiable. Queries with ambiguous instructions are minimally revised to enhance clarity
while preserving the original semantic intent. Furthermore, web-sourced queries lacking substantive
content are systematically eliminated. Crucially, all remaining queries undergo a final assessment of their
complexity and contextual relevance, ensuring only appropriately challenging and pertinent samples are
retained for the next stage.
Response Filtering. This phase integrates two complementary strategies:
• Rule-Based Filtering: A set of predefined heuristics is applied to eliminate responses exhibiting
qualitative deficiencies, such as repetition, incompleteness, or improper formatting. To maintain
semantic relevance and uphold ethical principles, we also discard any query-response pairs that are
off-topic or possess the potential to generate harmful content.
• Model-Based Filtering: The dataset is further refined by employing reward models derived from the
Qwen2.5-VL series. These models conduct a multi-dimensional evaluation of multimodal questionanswering pairs. Specifically: (a) answers are scored against a range of criteria, including correctness,
completeness, clarity, and helpfulness; (b) for vision-grounded tasks, the evaluation places special
emphasis on verifying the accurate interpretation and utilization of visual information; and (c) this
model-based approach enables the detection of subtle issues that typically elude rule-based methods,
such as inappropriate language mixing or abrupt stylistic shifts.
This multi-dimensional filtering framework ensures that only data meeting stringent criteria for quality,
reliability, and ethical integrity is advanced to the SFT phase.
4.2.2 Long-CoT Cold Start Data
The foundation of our thinking models is a meticulously curated Long Chain-of-Thought (CoT) cold
start dataset, engineered to elicit and refine complex reasoning capabilities. This dataset is built upon a
10
[Page 11 View] 1728x1216 (HxW)

[Page 11 Extracted Text]
diverse collection of queries spanning both pure-text and multimodal data, maintaining an approximate
1:1 ratio between vision-language and text-only samples to ensure balanced skill development.
The multimodal component, while covering established domains such as visual question answering
(VQA), optical character recognition (OCR), 2D/3D grounding, and video analysis, places a special
emphasis on enriching tasks related to STEM and agentic workflows. This strategic focus is designed to
push the model’s performance on problems requiring sophisticated, multi-step inference. The pure-text
portion closely mirrors the data used for Qwen3, featuring challenging problems in mathematics, code
generation, logical reasoning, and general STEM.
To guarantee high quality and an appropriate level of difficulty, we implement a rigorous multi-stage
filtering protocol.
• Difficulty Curation: We selectively retain instances where baseline models exhibited low pass rates or
generated longer, more detailed responses. This enriches the dataset with problems that are genuinely
challenging for current models.
• Multimodal Necessity Filtering: For vision-language mathematics problems, we introduce a critical
filtering step: we discard any samples that our Qwen3-30B-nothink model could solve correctly without
access to the visual input. This ensures that the remaining instances genuinely necessitate multimodal
understanding and are not solvable via textual cues alone.
• Response Quality Control: Aligning with the methodology of Qwen3, we sanitize the generated
responses. For queries with multiple candidate answers, we first remove those containing incorrect
final results. Subsequently, we filter out responses exhibiting undesirable patterns, such as excessive
repetition, improper language mixing, or answers that showed clear signs of guessing without sufficient
reasoning steps.
This stringent curation process yields a high-quality, challenging dataset tailored for bootstrapping
advanced multimodal reasoning.
4.3 Strong-to-Weak Distillation
We adopt the Strong-to-Weak Distillation pipeline as described in Qwen3 to further improve the performance of lightweight models. This distillation process consists of two main phases:
• Off-policy Distillation: In the first phase, outputs generated by teacher models are combined to
provide response distillation. This helps lightweight student models acquire fundamental reasoning
abilities, establishing a strong foundation for subsequent on-policy training.
• On-policy Distillation: In the second phase, the student model generates the responses based on the
provided prompts. These on-policy sequences are then used for fine-tuning the student model. We
align the logits predicted by the student and teacher by minimizing the KL divergence.
4.4 Reinforcement Learning
4.4.1 Reasoning Reinforcement Learning
We train models across a diverse set of text and multimodal tasks, including mathematics, coding, logical
reasoning, visual grounding, and visual puzzles. Each task is designed so that solutions can be verified
deterministically via rules or code executors.
Data Preparation We curate training data from both open-source and proprietary sources and apply
rigorous preprocessing and manual annotation to ensure high-quality RL queries. For multimodal
queries, we use a preliminary checkpoint of our most advanced vision–language model (Qwen3-VL-235BA22B) to sample 16 responses per query; any query for which all responses are incorrect is discarded.
We then run preliminary RL experiments per task to identify and remove data sources with limited
potential for improvement. This process yields approximately 30K RL queries covering a variety of text
and multimodal tasks. For training each model, we sample 16 responses for all queries and filter out
easy queries whose pass rate exceeds 90%. We shuffle and combine task-specific datasets to construct
mixed-task batches, ensuring a consistent, predefined ratio of samples per task. The ratio is determined
through extensive preliminary experiments.
Reward System We implement a unified reward framework that delivers precise feedback across all
tasks. The system provides shared infrastructure—data preprocessing, utility functions, and a reward
11
[Page 12 View] 1728x1216 (HxW)

[Page 12 Extracted Text]
manager to integrate multiple reward types—while the core reward logic is implemented per task. We
use task-specific format prompts to guide model outputs to the required formats and therefore do not rely
on explicit format rewards. To mitigate code-switching, we apply a penalty when the response language
differs from the prompt language.
RL Algorithm We employ SAPO (?), a smooth and adaptive policy-gradient method, for RL training.
SAPO delivers consistent improvements across diverse text and multimodal tasks and across different
model sizes and architectures.
4.4.2 General Reinforcement Learning
The General Reinforcement Learning (RL) stage is designed to enhance the model’s generalization
capabilities and operational robustness. To this end, we employ a multi-task RL paradigm where the
reward function is formulated based on a comprehensive set of tasks from the SFT phase, including VQA,
image captioning, OCR, document parsing, grounding, and clock recognition. The reward mechanism is
structured to optimize two principal dimensions of model performance:
• Instruction Following: This dimension evaluates the model’s adherence to explicit user directives. It
assesses the ability to handle complex constraints on content, format, length, and structured outputs
(e.g., JSON), ensuring the generated response precisely matches user requirements.
• Preference Alignment: For open-ended or subjective queries, this dimension aligns the model’s outputs
with human preferences by optimizing for helpfulness, factual accuracy, and stylistic appropriateness.
This fosters a more natural and engaging user interaction.
Furthermore, this stage acts as a corrective mechanism to unlearn strong but flawed knowledge priors
ingrained during SFT. We address this by introducing specialized, verifiable tasks designed to trigger
these specific errors, such as counter-intuitive object counting and complex clock time recognition. This
targeted intervention is designed to supplant erroneous priors with factual knowledge.
Another critical objective is to mitigate inferior behaviors like inappropriate language mixing, excessive
repetition, and formatting errors. However, the low prevalence of these issues makes general RL a
sample-inefficient correction strategy. To overcome this, we curate a dedicated dataset at this stage. This
dataset isolates prompts known to elicit such undesirable behaviors. This focused training enables the
application of targeted, high-frequency penalties, effectively suppressing these residual errors.
Feedback for the RL process is delivered via a hybrid reward system that combines two complementary
approaches:
• Rule-Based Rewards: This approach provides unambiguous, high-precision feedback for tasks with
verifiable ground truths, such as format adherence and instruction following. By using well-defined
heuristics, this method offers a robust mechanism for assessing correctness and effectively mitigates
reward hacking, where a model might exploit ambiguities in a learned reward function.
• Model-Based Rewards: This method employs Qwen2.5-VL-72B-Instruct or Qwen3 as sophisticated
judgers. The judge models evaluate each generated response against a ground-truth reference, scoring
its quality across multiple axes. This approach offers superior flexibility for assessing nuanced or openended tasks where strict, rule-based matching is inadequate. It is particularly effective at minimizing
false negatives that would otherwise penalize valid responses with unconventional formatting or
phrasing.
4.5 Thinking with Images
Inspired by the great prior works on "thinking with images" (????), we endow Qwen3-VL with similar
agentic capabilities through a two-stage training paradigm.
In the first stage, we synthesize a cold-start agentic dataset comprising approximately 10k grounding
examples—primarily simple two-turn visual question answering tasks such as attribute detection. We
then perform supervised fine-tuning (SFT) on Qwen2.5-VL-32B to emulate the behavior of a visual agent:
think → act → analyze feedback → answer. To further enhance its reasoning abilities, we apply multi-turn,
tool-integrated reinforcement learning (RL).
In the second stage, we distill the trained Qwen2.5-VL-32B visual agents from the first stage to generate a
larger, more diverse dataset of approximately 120k multi-turn agentic interactions spanning a broader
range of visual tasks. We then apply a similar cold-start SFT and tool-integrated RL pipeline (now using
both distilled and synthesized data) for the post-training of Qwen3-VL.
12
[PDF End]

{
  "file_path": "/path/to/qwen3vl.pdf",
  "pages": "13-18",
  "budget": "large"
}

[PDF Start]
Total pages: 35 | Showing pages: 13-18
[Page 13 View] 1728x1216 (HxW)

[Page 13 Extracted Text]
The multi-turn, tool-integrated RL procedure is nearly identical across both stages, differing only in
the underlying data. During RL, we employ three complementary reward signals to encourage robust,
tool-mediated reasoning:
• Answer Accuracy Reward leverages Qwen3-32B to measure whether the final answer is correct.
• Multi-Turn Reasoning Reward leverages Qwen2.5-VL-72B to evaluate whether the assistant correctly
interprets tool or environment feedback and arrives at the answer through coherent, step-by-step
reasoning.
• Tool-Calling Reward encourages appropriate tool usage by comparing the actual number of tool calls
to an expert-estimated target. This target is determined offline by Qwen2.5-VL-72B based on task
complexity.
Early experiments reveal a tendency for models to degenerate into making only a single tool call to
hack the first two rewards, regardless of task demands. To mitigate this, we explicitly incorporate the
tool-calling reward to promote adaptive tool exploration aligned with task complexity.
4.6 Infrastructure
We train the Qwen3-VL series models on Alibaba Cloud’s PAI-Lingjun AI Computing Service, which
provides the high-performance computing power required for compute-intensive scenarios such as AI
and high-performance computing.
During the pretraining phase, the system employs a hybrid parallelism strategy built upon the MegatronLM framework, integrating Tensor Parallelism (TP), Pipeline Parallelism (PP), Context Parallelism (CP),
Expert Parallelism (EP), and ZeRO-1 Data Parallelism (DP). This configuration achieves a fine-grained
balance among model scale, computational load, and communication overhead, enabling high hardware
utilization and sustaining both high throughput and low communication latency—even at scales of up to
10,000 GPUs.
For local deployment and performance evaluation, we adopt deployment strategies based on either vLLM
or SGLang. vLLM utilizes PagedAttention to enable memory-efficient management and high-throughput
inference, while SGLang excels at structured generation and handling complex prompts. Together, these
backends provide efficient inference and evaluation with stable, efficient, and flexible model inference
capabilities.
5 Evaluation
5.1 General Visual Question Answering
To comprehensively assess the general visual question answering (VQA) capabilities of the Qwen3-VL
series, we conduct extensive evaluations on a diverse set of benchmarks, including MMBench-V1.1 (?),
RealWorldQA (?), MMStar (?), and SimpleVQA (?). As detailed in Table 2, Table 3 and Table 4, the
Qwen3-VL family demonstrates robust and highly competitive performance across a wide spectrum of
model sizes, from 2B to 235B parameters.
In the comparison of thinking mode, Qwen3-VL-235B-A22B-Thinking achieves the highest score of 78.7 on
MMStar. Gemini-2.5-Pro’s (?) Thinking mode delivers the best overall performance, but Qwen3-VL-235BA22B-Thinking is not far behind. In the non-reasoning mode comparison, Qwen3-VL-235B-A22B-Instruct
obtains the highest scores on MMBench and RealWorldQA, with 89.3/88.9 and 79.2, respectively.
In the experiments with medium-sized models, Qwen3-VL-32B-Thinking achieves the highest scores on
MMBench and RealWorldQA, with 89.5/89.5 and 79.4, respectively. Notably, Qwen3-VL-32B-Instruct
even outperforms the Thinking variant on RealWorldQA, scoring 79.0.
The scalability of the Qwen3-VL series is evident in the strong performance of our smaller models.
Specifically, the largest model, Qwen3-VL-8B, achieves the highest performance across all five benchmarks.
For example, on MMBench-EN, the score in "thinking" mode increases from 79.9 for the 2B model to 85.3
for the 8B model. A similar upward trend is observed on other benchmarks, such as MMStar, where the
score rises from 68.1 (2B, thinking) to 75.3 (8B, thinking).
13
[Page 14 View] 1728x1216 (HxW)

[Page 14 Extracted Text]
5.2 Multimodal Reasoning
We evaluate the Qwen3-VL series on a wide range of multimodal reasoning benchmarks, primarily
focusing on STEM-related tasks and visual puzzles, including MMMU (?), MMMU-Pro (?), MathVision (?),
MathVision-Wildphoto (hereafter MathVisionWP) , MathVista (?), We-Math (?), MathVerse (?), DynaMath (?),
Math-VR (?), LogicVista (?), VisualPuzzles (?), VLM are Blind (?), ZeroBench (Main/Subtasks) (?), and
VisuLogic (?). As shown in Table 2, the flagship Qwen3-VL model demonstrates outstanding performance
across both “non-thinking” and “thinking” models. Notably, Qwen3-VL-235B-A22B-Instruct achieves
the best reported results among non-thinking or low-thinking-budget models on multiple benchmarks,
including MathVistamini, MathVision, MathVersemini, DynaMath, ZeroBench, VLMsAreBlind, VisuLogic,
and VisualPuzzlesDirect. While, Qwen3-VL-235B-A22B-Thinking achieves state-of-the-art results on
MathVistamini, MathVision, MathVersemini, ZeroBench, LogicVista, and VisuLogic.
Among medium-sized models, as shown in Table 3, Qwen3-VL-32B demonstrates significant advantages,
consistently outperforming Gemini-2.5-Flash and GPT-5-mini. Compared to the previous-generation
Qwen2.5-VL-72B model, the medium-sized Qwen3-VL model has already surpassed it on reasoning tasks.
This highlights significant progress in VLMs. Additionally, our newly introduced Qwen3-VL-30B-A3B
MoE model also delivers competitive results.
Among small-sized models, we compare Qwen3-VL-2B/4B/8B against GPT-5-Nano, with results presented in Table 4. The 8B variant maintains a clear advantage overall, while the 4B model achieves
the highest scores on DynaMath and VisuLogic. Notably, even the smallest 2B model exhibits strong
reasoning capabilities.
5.3 Alignment and Subjective Tasks
The ability to follow complex user instructions and reduce potential image-level hallucinations is indispensable for current large vision language models (VLMs). We assess our models on three representative
benchmarks: MM-MT-Bench (?), HallusionBench (?) and MIA-Bench (?). MM-MT-Bench is a multi-turn
LLM-as-a-judge evaluation benchmark for testing multimodal instruction-tuned models. HallusionBench
aims at diagnosing image-context reasoning and poses great challenges for current VLMs. MIA-Bench
is a more comprehensive benchmark to evaluate models’ reactions to users’ complex instructions (e.g.,
creative writing with character limit and compositional instructions).
As shown in Table 2, our flagship Qwen3-VL-235B-A22B model consistently outperforms other closedsource models. On HallusionBench, our thinking version surpasses Gemini-2.5-pro (?), GPT-5 (?) and
Claude opus 4.1 (?) by 3.0, 1.0, and 6.3 points, respectively. On MIA-Bench, Qwen3-VL-235B-A22BThinking achieves the overall best score across all the other models, showing our superior multimodal
instruction following ability. We also investigate detailed subtask results of MIA-Bench: our model
overtakes GPT-5-high-thinking version by 10.0 and 5.0 points in math and textual subtasks of MIA-Bench,
respectively. The same trend can be observed on our smaller-sized models like Qwen3-VL-30B-A3B, and
Qwen3-VL-32B, where they overtake other models with comparable sizes. Our 2B/4B/8B series also
performs well and shows a negligible drop, especially on MIA-Bench.
5.4 Text Recognition and Document Understanding
We compare the Qwen3-VL series with other models of comparable size on document-related benchmarks,
including OCR, document parsing, document question answering (QA), and document reasoning.
We evaluate our flagship model, Qwen3-VL-235B-A22B, against state-of-the-art VLMs on the benchmarks
listed in Table 2. On OCR-focused parsing benchmarks — including CC-OCR (?) and OmniDocBench (?)
— as well as comprehensive OCR benchmarks such as OCRBench (?) and OCRBench_v2 (?), the Qwen3-
VL-235B-A22B-Instruct model establishes a new state of the art, marginally outperforming its “thinking”
counterpart, Qwen3-VL-235B-A22B-Thinking. On OCR-related visual question answering (VQA) benchmarks that require both OCR capability and keyword search — such as DocVQA (?), InfoVQA (?),
AI2D (?), ChartQA (?), and the CharXiv (?) description subset — both the Instruct and Thinking variants
achieve comparable performance, demonstrating consistently strong results across these tasks. Notably,
on the reasoning subset of CharXiv — which demands deep chart comprehension and multi-step reasoning — the Thinking variant surpasses the Instruct version and ranks second only to GPT5-thinking and
Gemini-2.5-Pro-Thinking.
Furthermore, among the smaller-sized variants in the Qwen3-VL series, both Qwen3-VL-30BA3B models
and Qwen3-VL-32B models consistently outperform Gemini-2.5-Flash and GPT-5-mini across most
evaluation metrics, as shown in Table 3. Even the compact dense models — Qwen3-VL-8B, Qwen3-VL-4B,
and Qwen3-VL-2B — demonstrate remarkably competitive performance on OCR parsing, visual question
14
[Page 15 View] 1728x1216 (HxW)

[Page 15 Extracted Text]
Table 2: Performance of Qwen3-VL-235B-A22B and top-tier models on visual benchmarks. The highest
scores of the reasoning and non-reasoning models are shown in bold and underlined, respectively.
Results marked with an ∗ are sourced from the technical report. + denotes results with tool use.
Benchmark
thinking instruct thinking budget-128 high minimal thinking non-thinking
STEM
Puzzle
MMMU 80.6 78.7 81.7∗ 80.9 84.2
∗ 74.4∗ 78.4 77.2
MMMU-Pro 69.3 68.1 68.8∗ 71.2 78.4
∗ 62.7∗ 64.8 60.7
MathVistamini 85.8 84.9 82.7∗ 77.7 81.3 50.9 75.5 74.5
MathVision 74.6 66.5 73.3∗ 66.0 70.9 45.8 64.3 57.7
MathVisionWP 63.8 57.0 63.2 56.9 62.8 40.1 54.0 46.4
We-Math 74.8 67.5 80.6 74.5 73.8 51.8 65.2 60.2
MathVersemini 85.0 72.5 82.9 65.9 84.1 43.0 70.6 68.1
DynaMath 82.8 79.4 80.0 78.5 85.4 74.0 75.1 72.0
Math-VR 66.8 65.0 64.7* 54.3 58.1 21.7 54.3 38.0
ZeroBench 4 2 3 1 2 2 3 1
VlmsAreBlind 79.5 80.4 86.1 78.5 80.5 53.4 77.8 72.2
LogicVista 72.2 65.8 72.0 68.7 71.8 46.3 67.3 63.5
VisuLogic 34.4 29.9 31.6 26.9 28.5 27.2 27.9 27.2
VisualPuzzles 57.2 54.7 60.9 56.9 57.3 47.9 48.8 47.6
General VQA
MMBench-EN 88.8 89.3 90.1
∗ 88.4 83.8 81.3 79.4 83.0
MMBench-CN 88.6 88.9 89.7
∗ 86.4 83.5 79.9 84.9 74.3
RealWorldQA 81.3 79.2 78.0∗ 76.0 82.8 77.3 69.9 68.5
MMStar 78.7 78.4 77.5∗ 78.5 76.4 65.2 72.1 71.0
SimpleVQA 61.3 63.0 65.4 66.9 61.8 56.7 56.7 55.7
Alignment
HallusionBench 66.7 63.2 63.7∗ 60.9 65.7 53.7 60.4 55.1
MM-MT-Bench 8.5 8.5 8.4∗ 7.6 7.6 7.5 7.8 7.9
MIA-Bench 92.7 91.3 92.3 91.3 92.4 92.6 91.2 90.0
Document
Understanding
DocVQAtest 96.5 97.1 92.6 94.0 91.5 89.6 92.5 89.2
InfoVQAtest 89.5 89.2 84.2 82.9 79.0 69.9 69.4 60.9
AI2Dw. M. 89.2 89.7 90.9 90.0 89.7 84.1 86.4 84.4
ChartQAtest 90.3 90.3 83.3 62.6 59.7 59.1 86.2 83.9
OCRBench 875 920 866 872 810 787 764 750
OCRBench_v2en 66.8 67.1 54.3 55.2 53.0 48.2 48.4 47.2
OCRBench_v2zh 63.5 61.8 48.5 53.1 43.2 37.7 43.7 38.0
CC-OCR 81.5 82.2 77.2 76.8 68.3 66.1 69.1 66.0
OmniDocBenchen 0.155 0.143 0.347 0.206 0.356 0.174 0.194 -
OmniDocBenchzh 0.207 0.207 0.238 0.249 0.472 0.389 0.293 -
CharXiv(DQ) 90.5 89.4 94.4 87.8 89.2 79.5 88.5 87.8
CharXiv(RQ) 66.1 62.1 67.9 62.9 81.1
∗ 57.8 63.6 60.2
MMLongBenchDoc 56.2 57.0 55.6 51.2 51.5 42.4 54.5 48.1
2D/3D
Grounding
RefCOCO-avg 92.1 91.9 74.6∗- 66.8 - - -
CountBench 93.7 93.0 91.0∗ 91.0 91.7 87.8 93.1 91.9
ODinW-13 43.2 48.6 33.7∗ 34.5 - - - -
ARKitScenes 53.7 56.9 - - - - - -
Hypersim 11.0 13.0 - - - - - -
SUNRGBD 34.9 39.4 29.7 - - - - -
Embodied/Spatial
Understanding
ERQA 52.5 51.3 55.3 50.3 65.7∗ 42.0∗ 34.8 28.0
VSI-Bench 60.0 62.7 - - - - - -
EmbSpatialBench 84.3 83.1 79.1 73.3 82.9 75.1 69.2 66.0
RefSpatialBench 69.9 65.5 36.5 35.6 23.8 23.1 - -
RoboSpatialHome 73.9 69.4 47.5 49.2 53.5 43.6 - -
Multi-Image BLINK 67.1 70.7 70.6∗ 70.0 71.0 62.8 64.1 62.9
MUIRBENCH 80.1 73.0 77.2 74.0 77.5 66.5 - -
Video
Understanding
MVBench 75.2 76.5 69.9 65.8 75.3 64.6 61.4 59.0
Video-MMEw/o sub. 79.0 79.2 85.1 80.6 84.7 77.3 75.6 73.3
MLVUM-Avg 83.8 84.3 85.6 81.2 86.2 78.3 73.5 71.2
LVBench 63.6 67.7 73.0 69.0 - - - -
Charades-STAmIoU 63.5 64.8 - - - - - -
VideoMMMU 80.0 74.7 83.6∗ 79.4 84.6∗ 61.6∗ 76.2 70.1
MMVU 71.1 68.1 74.9 72.2 73.0 68.1 66.4 61.4
Perception
with Tool
V
∗ 85.9 93.7+ 83.8 72.7 72.8 56.7 - -
HRBench4K 84.3 85.4+ 87.3 84.8 - - - -
HRBench8K 76.6 82.4+ 85.4 80.1 - - - -
Multi-Modal
Coding
Design2Code 93.4 92.0 89.2 90.3 92.5 88.9 88.5 85.3
ChartMimic 78.4 80.5 83.9 79.9 62.1 41.4 85.2 82.9
UniSVG 65.8 69.8 70.0 67.9 71.7 74.5 73.0 72.5
Multi-Modal
Agent
ScreenSpot Pro 61.8 62.0 - - - - - -
OSWorldG 68.3 66.7 45.2 - - - - -
AndroidWorld 62.0 63.7 - - - - - -
OSWorld 38.1 31.6 - - - - - 44.4
WindowsAA 32.1 28.9 - - - - - -
Qwen3-VL
235B-A22B
Gemini
2.5 Pro
OpenAI
GPT-5
Claude
Opus 4.1
15
[Page 16 View] 1728x1216 (HxW)

[Page 16 Extracted Text]
Table 3: Performance of medium-sized Qwen3-VL models and previous models on visual benchmarks.
The highest scores are shown in bold. Results marked with an ∗ are sourced from the technical report. +
denotes results with tool use.
Benchmark
thinking instruct thinking instruct thinking non-thinking high minimal
STEM
Puzzle
MMMU 76.0 74.2 78.1 76.0 77.7 76.3 79.0 67.9
MMMU-Pro 63.0 60.4 68.1 65.3 67.2 65.9 67.3 53.7
MathVistamini 81.9 80.1 85.9 83.8 79.4 75.3 79.1 59.6
MathVision 65.7 60.2 70.2 63.4 64.3 60.7 71.9 46.6
MathVisionWP 58.9 52.3 58.6 54.6 53.6 49.0 56.6 42.8
We-Math 70.0 56.9 71.6 63.3 53.9 60.3 70.2 51.4
MathVersemini 79.6 70.2 82.6 76.8 77.7 75.9 78.8 36.5
DynaMath 80.1 73.4 82.0 76.7 75.9 69.7 81.4 71.3
Math-VR 61.7 61.3 62.3 59.8 58.8 54.7 58.2 26.4
ZeroBench 0 0 2 1 1 3 3 2
VlmsAreBlind 72.5 67.5 85.1 87.0 77.5 75.9 75.8 62.0
LogicVista 65.8 53.5 70.9 62.2 67.3 60.0 71.4 50.8
VisuLogic 26.6 23.0 32.4 29.7 31.0 23.3 27.2 27.6
VisualPuzzles 52.0 46.2 54.7 53.2 41.4 45.0 59.3 48.2
General VQA
MMBench-EN 87.0 86.1 89.5 87.6 87.1 86.6 86.6 78.5
MMBench-CN 85.9 85.3 89.4 87.7 87.3 86.0 84.0 76.3
RealWorldQA 77.4 73.7 78.4 79.0 76.0 75.7 79.0 73.3
MMStar 75.5 72.1 79.4 77.7 76.5 75.8 74.1 61.3
SimpleVQA 54.3 52.7 55.4 56.9 63.2 59.2 56.8 50.3
Alignment
HallusionBench 66.0 61.5 67.4 63.8 63.5 59.1 63.2 55.9
MM-MT-Bench 7.9 8.0 8.3 8.4 8.1 8.0 7.7 7.4
MIA-Bench 91.6 91.2 92.3 91.8 91.1 90.6 92.0 92.3
Document
Understanding
DocVQAtest 95.5 95.0 96.1 96.9 92.8 93.0 90.5 90.6
InfoVQAtest 85.6 81.8 89.2 87.0 82.5 81.7 77.6 72.8
AI2Dw. M. 86.9 85.0 88.9 89.5 88.7 87.7 88.2 82.9
ChartQAtest 89.4 86.8 89.0 88.5 60.6 69.0 57.5 57.8
OCRBench 839 903 855 895 853 864 821 807
OCRBench_v2en 62.6 63.2 68.4 67.4 52.2 50.6 52.6 45.7
OCRBench_v2zh 60.4 57.8 62.1 59.2 43.8 43.9 45.1 41.0
CC-OCR 77.8 80.7 79.6 80.3 75.4 74.8 70.8 61.6
OmniDocBenchen 0.165 0.183 0.148 0.151 0.265 0.228 0.181 0.260
OmniDocBenchzh 0.233 0.253 0.236 0.239 0.245 0.305 0.316 0.425
CharXiv(DQ) 86.9 85.5 90.2 90.5 90.1 85.5 89.4 78.6
CharXiv(RQ) 56.6 48.9 65.2 62.8 61.7 60.1 68.6 48.9
MMLongBenchDoc 47.4 47.1 54.6 55.4 49.0 44.6 50.3 39.6
2D/3D
Grounding
RefCOCO-avg 89.3 89.7 91.1 91.9 - - - -
CountBench 90.0 89.8 94.1 94.9 86.0 83.7 91.0 84.1
ODinW-13 42.3 47.5 41.8 46.6 - - - -
ARKitScenes 55.6 56.1 46.1 55.6 - - - -
Hypersim 11.4 12.5 12.5 14.0 - - - -
SUNRGBD 34.6 38.1 33.9 37.0 - - - -
Embodied/Spatial
Understanding
ERQA 45.3 43.0 52.3 48.8 - - 54.0 45.8
VSI-Bench 56.1 63.2 61.2 61.5 - - 31.5 30.5
EmbSpatialBench 80.6 76.4 82.7 81.5 - - 80.7 72.1
RefSpatialBench 54.2 53.1 67.2 61.4 - - 9.0 4.0
RoboSpatialHome 65.5 62.9 74.2 64.6 - - 54.3 44.6
Multi-Image BLINK 65.4 67.7 68.5 67.3 68.1 66.8 - 56.7
MUIRBENCH 77.6 62.9 80.3 72.8 72.7 67.5 - 57.5
Video
Understanding
MVBench 72.0 72.3 73.2 72.8 - - - -
Video-MMEw/o sub. 73.3 74.5 77.3 76.6 79.6 75.6 78.9 71.0
MLVUM-Avg 78.9 81.3 82.3 82.1 82.1 77.8 83.3 71.7
LVBench 59.2 62.5 62.6 63.8 64.5 62.2 - -
Charades-STAmIoU 62.7 63.5 62.8 61.2 - - - -
VideoMMMU 75.0 68.7 79.0 71.9 73.9 65.2 82.5* 56.7
MMVU 66.1 59.8 67.9 66.8 69.8 68.2 69.8 64.8
Perception
with Tool
V
∗ 81.2 89.5+ 84.8 91.1+ - - 78.6 63.9
HRBench4K 77.8 82.5+ 82.1 84.6+ - - 78.6 66.3
HRBench8K 71.3 79.3+ 74.8 81.6+ - - 74.4 60.9
Multi-Modal
Agent
ScreenSpot Pro 57.3 60.5 57.1 57.9 - - - -
OSWorldG 59.6 61.0 64.0 65.1 - - - -
AndroidWorld 55.0 54.3 63.7 57.3 - - - -
OSWorld 30.6 30.3 41.0 32.6 - - - -
WindowsAA 24.2 24.9 42.9 30.9 - - - -
Qwen3-VL
30B-A3B
Qwen3-VL
32B
Gemini
2.5 Flash
GPT-5
mini
16
[Page 17 View] 1728x1216 (HxW)

[Page 17 Extracted Text]
Figure 2: Multilingual OCR performance of our model on a self-built test set. The model achieves
over 70% accuracy on 32 out of 39 supported languages, demonstrating strong and usable multilingual
capabilities.
answering (VQA), and comprehensive benchmark suites, as detailed in Table 4. This highlights the
exceptional efficiency and strong scalability of the Qwen3-VL architecture across model sizes.
In this version of the Qwen3-VL, we have placed particular emphasis on enhancing its ability to understand long documents. As reported in Table 2, in the comparison within the flagship models on the
MMLongBench-Doc benchmark (?), our Qwen3-VL-235B-A22B achieves overall accuracy of 57.0%/56.2%
under the instruct/thinking settings, showcasing the SOTA performance on the long document understanding task.
Beyond its strong performance on established benchmarks, we have also made substantial strides in
multilingual support. This represents a major expansion from the 10 non-English/Chinese languages
supported by Qwen2.5-VL to 39 languages in Qwen3-VL. We assess this expanded capability on a newly
constructed, in-house dataset. As illustrated in Figure 2, the model’s accuracy surpasses 70%—a threshold
we consider practical for real-world usability—on 32 out of the 39 languages tested. This demonstrates
that the strong OCR capabilities of Qwen3-VL are not confined to a handful of languages but extend
across a broad and diverse linguistic spectrum.
5.5 2D and 3D Grounding
In this section, we conduct a comprehensive evaluation of the Qwen3-VL series on both 2D and 3D
grounding-related benchmarks and compare the models with state-of-the-art models that possess similar
capabilities.
We evaluate Qwen3-VL’s 2D grounding capabilities on the referring expression comprehension benchmarks RefCOCO/+/g (??), the open-vocabulary object detection benchmark ODinW-13 (?), and the
counting benchmark CountBench (?). For ODinW-13, we adopt mean Average Precision (mAP) as the
evaluation metric by setting confidence scores to 1.0. To ensure comparability with conventional open-set
object detection specialist models, we provide all dataset categories simultaneously within the prompt
during evaluation. As shown in Table 2, our flagship model, Qwen3-VL-235B-A22B, demonstrates
outstanding performance and achieves state-of-the-art (SOTA) results across 2D grounding and counting
benchmarks. Notably, it achieves 48.6 mAP on ODinW-13, demonstrating strong performance in multitarget open-vocabulary object grounding. Detailed results for our smaller-scale variants, which also
exhibit competitive performance in 2D visual grounding, are presented in Tables 3 and 4, respectively.
Moreover, in this version of Qwen3-VL, we enhance its spatial perception capabilities for 3D object
localization. We evaluate the Qwen3-VL series against other models of comparable scale on Omni3D (?),
a comprehensive benchmark comprising datasets such as ARKitScenes (?), Hypersim (?), and SUN
RGB-D (?). We employ mean Average Precision (mAP) as our evaluation metric. Each input is an imagetext pair consisting of the image and a textual prompt specifying the object category. To ensure a fair
17
[Page 18 View] 1728x1216 (HxW)

[Page 18 Extracted Text]
Table 4: Performance of small-sized Qwen3-VL models and GPT-5-nano on visual benchmarks.
Benchmark
thinking instruct thinking instruct thinking instruct high minimal
STEM
Puzzle
MMMU 61.4 53.4 70.8 67.4 74.1 69.6 75.8 57.6
MMMU-Pro 42.5 36.5 57.0 53.2 60.4 55.9 57.2 36.5
MathVistamini 73.6 61.3 79.5 73.7 81.4 77.2 71.5 40.9
MathVision 45.9 31.6 60.0 51.6 62.7 53.9 62.2 33.2
MathVisionWP 35.5 30.9 48.7 44.4 53.3 45.4 49.3 28.3
MathVersemini 66.9 52.1 75.2 46.8 77.7 62.1 74.2 27.0
DynaMath 66.7 54.2 74.4 65.3 73.2 67.7 78.0 62.0
Math-VR 37.7 20.7 58.1 52.3 59.0 53.4 49.7 25.0
ZeroBench 0 0 0 0 2 1 1 1
VlmsAreBlind 50.0 56.0 68.6 71.9 69.1 74.0 66.7 40.2
LogicVista 50.0 35.8 61.1 53.2 65.1 55.3 59.7 40.5
VisuLogic 25.4 11.5 30.2 19.0 27.5 22.5 24.5 24.0
VisualPuzzles 37.4 34.3 48.9 43.7 51.7 47.9 43.5 31.3
General VQA
MMBench-EN 79.9 78.4 84.6 83.9 85.3 84.5 78.4 50.8
MMBench-CN 78.8 75.9 83.8 83.5 85.5 84.7 77.6 48.5
RealWorldQA 69.5 63.9 73.2 70.9 73.5 71.5 71.8 60.7
MMStar 68.1 58.3 73.2 69.8 75.3 70.9 68.6 41.3
SimpleVQA 43.6 40.7 48.8 48.0 49.6 50.2 46.0 39.0
Alignment
HallusionBench 54.9 51.4 64.1 57.6 65.4 61.1 58.4 39.3
MM-MT-Bench 6.9 5.9 7.7 7.5 8.0 7.7 6.6 6.2
MIA-Bench 85.6 83.6 91.0 89.7 91.5 91.1 89.9 89.6
Document
Understanding
DocVQAtest 92.9 93.3 94.2 95.3 95.3 96.1 88.2 78.3
InfoVQAtest 77.1 72.4 83.0 80.3 86.0 83.1 68.6 49.2
AI2Dw. M. 80.4 76.9 84.9 84.1 84.9 85.7 81.9 65.7
ChartQAtest 86.6 79.1 88.8 84.6 88.6 89.6 52.1 48.6
OCRBench 792 858 808 881 819 896 753 701
OCRBench_v2en 56.4 56.3 61.8 63.7 63.9 65.4 48.1 37.9
OCRBench_v2zh 51.9 53.0 55.8 57.6 59.2 61.2 33.6 27.3
CC-OCR 68.3 72.8 73.8 76.2 76.3 79.9 58.9 52.9
OmniDocBenchen 0.370 0.292 0.234 0.244 0.209 0.170 0.401 0.454
OmniDocBenchzh 0.447 0.348 0.297 0.285 0.253 0.264 0.518 0.568
CharXiv(DQ) 70.1 62.3 83.9 76.2 85.9 83.0 82.0 64.4
CharXiv(RQ) 37.1 26.8 50.3 39.7 53.0 46.4 50.1 31.7
MMLongBenchDoc 33.8 31.6 44.4 43.5 48.0 47.9 31.8 22.1
2D/3D
Grounding
RefCOCO-avg 84.8 85.6 88.2 89.0 88.2 89.1 - -
CountBench 84.1 88.4 89.4 84.9 91.5 80.5 80.0 62.9
ODinW-13 36.0 43.4 39.4 48.2 39.8 44.7 - -
ARKitScenes 47.7 56.2 46.3 56.6 46.6 56.8 - -
Hypersim 11.2 12.0 11.9 12.2 12.0 12.7 - -
SUNRGBD 28.6 33.8 28.0 34.7 30.4 36.2 - -
Embodied/Spatial
Understanding
ERQA 41.8 28.3 47.3 41.3 46.8 45.8 45.8 37.8
VSI-Bench 48.0 53.9 55.2 59.3 56.6 59.4 15.4 27.0
EmbSpatialBench 75.9 69.2 80.7 79.6 81.1 78.5 74.2 50.7
RefSpatialBench 28.9 30.3 45.3 46.6 44.6 54.2 12.6 2.5
RoboSpatialHome 45.3 49.1 63.2 61.7 62.0 66.9 46.1 44.8
Multi-Image BLINK 57.2 53.8 63.4 65.8 64.7 69.1 58.3 42.2
MUIRBENCH 68.1 47.4 75.0 63.8 76.8 64.4 65.7 45.7
Video
Understanding
MVBench 64.5 61.7 69.3 68.9 69.0 68.7 - -
Video-MMEw/o sub. 62.1 61.9 68.9 69.3 71.8 71.4 66.2 49.4
MLVUM-Avg 69.2 68.3 75.7 75.3 75.1 78.1 69.2 52.6
LVBench 47.6 47.4 53.5 56.2 55.8 58.0 - -
Charades-STAmIoU 56.9 54.5 59.0 55.5 59.9 56.0 - -
VideoMMMU 54.1 41.9 69.4 56.2 72.8 65.3 63.0 40.2
MMVU 48.9 41.7 58.6 50.5 62.0 58.7 63.1 51.0
Perception
with Tool
V
∗ 69.1 75.9+ 74.9 88.0+ 77.5 90.1+ - -
HRBench4K 69.4 72.6+ 73.5 81.3+ 72.4 82.3+ - -
HRBench8K 62.6 68.9+ 67.1 74.4+ 68.1 78.0+ - -
Multi-Modal
Agent
ScreenSpot Pro 32.2 48.5 49.2 59.5 46.6 54.6 - -
OSWorldG 41.8 46.1 53.9 58.2 56.7 58.2 - -
AndroidWorld 46.1 36.4 52.0 45.3 50.0 47.6 - -
OSWorld 19.0 17.0 31.4 26.2 33.9 33.9 - -
WindowsAA - - 35.5 23.4 24.1 28.8 - -
Qwen3-VL
2B
Qwen3-VL
4B
Qwen3-VL
8B
OpenAI
GPT-5 nano
18
[PDF End]

{
  "file_path": "/path/to/qwen3vl.pdf",
  "pages": "17",
  "dpi": 300,
  "output_dir": "/tmp"
}

Rendered 1/35 page(s) of /path/to/qwen3vl.pdf @ 300 DPI:
  [page 17] 3508x2481 -> /path/to/qwen3vl_page_17.png
[page 17] 1216x864 (HxW)

{
  "image_path": "/path/to/qwen3vl_page_17.png",
  "box": [
    150,
    85,
    910,
    360
  ],
  "output_path": "/path/to/fig2_crop.png"
}

Cropped: /path/to/qwen3vl_page_17.png
Box: [150,85,910,360] → pixel [372,298,2258,1263]
Saved to: /path/to/fig2_crop.png | 965x1886 (HxW)
[Preview] 736x1440 (HxW)