Qwen VL Plus: Enhanced Vision-Language Models

• Qwen VL Plus is a family of enhanced vision-language models characterized by a unified architecture that integrates visual and textual processing through a three-stage training pipeline.
• Key architectural upgrades include advanced visual encoders, position-aware adapters, and long-context integration that support tasks spanning image, video, OCR, and retrieval.
• Recent research demonstrates the model’s adaptability via parameter-efficient tuning and specialized descendants for retrieval, gaze-conditioned processing, and embodied control.

Searching arXiv for papers relevant to Qwen VL Plus and related Qwen-VL model lines. In several papers, Qwen VL Plus is used for Qwen2.5-VL-7B and for the broader Qwen VL Plus family. This suggests a practical usage in which the term denotes enhanced Qwen vision-LLMs rather than a single, formally fixed architecture. Within that usage, the line spans the original Qwen-VL models built on Qwen-7B, later long-context systems such as Qwen3-VL, and specialized descendants for retrieval, gaze-conditioned video understanding, and embodied control (Bai et al., 2023, Bai et al., 26 Nov 2025, Li et al., 8 Jan 2026, Pham et al., 26 Mar 2026, Wang et al., 28 May 2026).

1. Terminology, scope, and model lineage

The foundational reference point is "Qwen-VL: A Versatile Vision-LLM for Understanding, Localization, Text Reading, and Beyond" (Bai et al., 2023). That work introduces the Qwen-VL series, a set of large-scale vision-LLMs designed to perceive and understand both texts and images, starting from the Qwen-7B LLM and adding visual capacity through a visual receptor, an input-output interface, a 3-stage training pipeline, and a multilingual multimodal cleaned corpus. The resulting models include Qwen-VL and Qwen-VL-Chat, with the latter instruction-tuned for dialogue.

Later reports extend the same line substantially. Qwen3-VL is presented as the most capable vision-LLM in the Qwen series to date, with dense variants at 2B/4B/8B/32B and MoE variants at 30B-A3B/235B-A22B, and native support for interleaved contexts of up to 256K tokens across text, images, and video (Bai et al., 26 Nov 2025). Further specialization yields Qwen3-VL-Embedding and Qwen3-VL-Reranker, which adapt the foundation into a unified retrieval and ranking pipeline (Li et al., 8 Jan 2026). In a different direction, Qwen-VLA extends the Qwen vision-language stack from perception, understanding, and reasoning to continuous action and trajectory generation through a DiT-based action decoder (Wang et al., 28 May 2026).

Because several later papers explicitly gloss Qwen2.5-VL-7B as Qwen VL Plus, the term functions in the literature as a convenient handle for a mature stage of the Qwen multimodal line rather than as a uniquely delimited model card. A plausible implication is that discussions of “Qwen VL Plus” should be read against the evolving Qwen-VL architecture family, not as references to a single invariant system.

2. Core architecture and training stack

The original Qwen-VL architecture couples a ViT-bigG from OpenCLIP visual encoder with a position-aware vision-language adapter and the Qwen-7B LLM (Bai et al., 2023). The adapter is a single-layer cross-attention module that compresses patchwise visual outputs into a fixed-length sequence of 256 tokens using trainable query embeddings and 2D absolute positional encodings. Images are marked by special tokens such as <img> and </img>, while grounding tasks serialize normalized bounding boxes as "<box>(X_{tl},Y_{tl}),(X_{br},Y_{br})</box>" and use <ref> tags for text-region alignment.

Its training is organized into three stages. Stage 1 performs pre-training on 1.4 billion cleaned image-text pairs, with the LLM frozen and images at 224×224. Stage 2 performs multi-task pre-training at 448×448 on interleaved image-text sequences up to length 2048, covering captioning, VQA, grounding, OCR, and related tasks. Stage 3 performs supervised fine-tuning for Qwen-VL-Chat on up to 350K dialogue examples, with the visual encoder frozen. The corpus is explicitly multilingual, with the Stage 1 mixture described as 77.3% English and 22.7% Chinese.

Qwen3-VL keeps the unified multimodal orientation but changes the technical envelope materially (Bai et al., 26 Nov 2025). The report identifies three major upgrades: interleaved-MRoPE for spatial-temporal modeling, DeepStack integration for multi-level ViT feature injection, and text-based time alignment for video. The family uses a transformer-based LLM, a vision encoder with SigLIP-2 as the default, and an MLP merger for vision-language fusion. The training stack also adds token budgeting, square-root loss reweighting, and strong-to-weak distillation. This progression marks a shift from an adapter-centric image-text design toward a long-context, interleaved multimodal backbone.

The remote-sensing Change VQA study further sharpens the architectural distinction between structured and native multimodality (Bazi et al., 20 Apr 2026). Under a unified LoRA setting, it compares Qwen3-VL, which uses a structured vision-language pipeline with multi-depth visual conditioning and a full self-attention backbone, against Qwen3.5, described there as a native multimodal model with single-stage alignment and a hybrid backbone combining GatedDeltaNet blocks and full-attention layers. That comparison is important because it isolates architecture under matched adaptation conditions rather than conflating it with full-model retraining.

3. Capability profile across understanding, grounding, OCR, and long context

Qwen-VL was introduced as a generalist model rather than a narrowly specialized one, and its benchmark profile reflects that ambition (Bai et al., 2023). On image captioning, the report gives 85.8 CIDEr on Flickr30K in the zero-shot setting. For general VQA, Qwen-VL-Chat is reported at 78.2 on VQAv2, 56.6 on OKVQA, 57.5 on GQA, 68.2 on ScienceQA-Img, and 38.9 on VizWiz. On text-oriented tasks, Qwen-VL reaches 63.8 on TextVQA and also leads on DocVQA, ChartQA, AI2D, and OCR-VQA. For grounding, Qwen-VL-7B is reported at 89.4–92.3% on RefCOCO val/testA, with substantial gains over earlier open-source generalist models. For real-world dialogue, Qwen-VL-Chat attains 645.2 (EN) and 401.2 (CN) on TouchStone, 65.4 on image and 37.8 on video for SEED-Bench, and 1487.58/360.71 on MME (Perception/Cognition).

Qwen3-VL expands this profile into long-context reasoning and video (Bai et al., 26 Nov 2025). The report attributes three core pillars to the family: stronger pure-text understanding, robust long-context comprehension with a native 256K-token window for interleaved multimodal inputs, and advanced multimodal reasoning across single-image, multi-image, and video tasks. Representative numbers include 85.8% on MathVista (mini) for Qwen3-VL-235B-A22B-Thinking, 74.6% on MathVision, and up to 80.6% on MMMU in thinking mode. For long-form video retrieval-style evaluation, the report states >99.5% recall at up to 1M tokens (~2 hours video) on needle-in-the-haystack evaluations.

The technical significance of this benchmark spread is that the Qwen line is not confined to generic captioning or visual chat. The family is repeatedly evaluated on OCR-rich documents, visual grounding, multi-image reasoning, and long-horizon video comprehension. This breadth helps explain why later papers reuse Qwen backbones for retrieval, agentic interfaces, remote sensing, and robotics rather than treating them as single-purpose VQA systems.

4. Adaptation, retrieval, and specialized multimodal reasoning

A recurring theme in the literature is that Qwen VL Plus-class models remain highly adaptable under parameter-efficient tuning. In the Change VQA remote-sensing study, all models are fine-tuned with LoRA applied only to decoder attention projections—query/key/value/output—with vision encoders and alignment modules frozen, using rank $r=16$ and scale $\alpha=32$ (Bazi et al., 20 Apr 2026). The update is written as

$$\mathbf{W} = \mathbf{W}_0 + \frac{\alpha}{r}\mathbf{B}\mathbf{A}.$$

Under this setting, Qwen3.5-2B improves OA on Test2 from 65.38 for Qwen3-VL-2B to 70.94, and reaches AA/OA: 69.56/70.94, exceeding the cited best prior VisTA score of 65.9/68.5. The paper’s central conclusion is that performance does not scale monotonically with model size and that native multimodal models are more effective than structured vision-language pipelines for this task.

The retrieval branch makes the specialization explicit. Qwen3-VL-Embedding is a bi-encoder that maps text, images, document images, and video into a unified semantic space, while Qwen3-VL-Reranker is a cross-encoder that performs fine-grained relevance estimation with cross-attention (Li et al., 8 Jan 2026). Both support 32k tokens, both are released in 2B and 8B variants, and the embedding model adds Matryoshka Representation Learning and Quantization-Aware Training. The report states that Qwen3-VL-Embedding-8B attains an overall 77.8 on MMEB-V2, ranking first among all models as of January 8, 2025, while also reaching 67.9 on MMTEB.

Another specialization route is architectural modulation rather than retraining the backbone. GazeQwen equips Qwen2.5-VL-7B with a compact gaze-conditioned resampler of roughly 1–5M trainable parameters, with an optional second stage adding LoRA of about 3.5M parameters, for a total of roughly 5–9M trainable parameters (Pham et al., 26 Mar 2026). On the StreamGaze benchmark it reports 63.9% mean accuracy, a +16.1 point gain over the same backbone with gaze as visual prompts and +10.5 points over GPT-4o. The paper states that learning where to inject gaze within an LLM is more effective than scaling model size or engineering better prompts. In the context of Qwen VL Plus, this is evidence that hidden-state level integration can be more important than input-format heuristics.

5. Agentic web interaction and embodied extensions

The literature gives a mixed picture for agentic use. On the web-browsing benchmark MMSearch-Plus, Qwen-2.5-VL-72B-Instruct records 0.0% accuracy without search, 13.5% with one round of image search, and 6.9% under full rollout with up to 10 rounds of image/text search (Tao et al., 29 Aug 2025). The same report notes 421 invalid tool calls across 311 tasks, describes failures in source verification, part-based reasoning, bounding-box production, and long-horizon planning, and states that enabling agent-driven cropping frequently deteriorated performance. It also describes Qwen as the strongest open-source model tested, but still far behind the strongest closed-source agent under that framework.

Embodied extensions use the Qwen multimodal stack more successfully when the task is formulated as perception-to-action rather than open-web browsing. Qwen-VLA uses Qwen3.5-4B as an early-fusion vision-language backbone and adds a DiT-based action decoder of roughly 1.15B parameters together with embodiment-aware prompt conditioning (Wang et al., 28 May 2026). Manipulation, navigation, and trajectory prediction are cast into a unified action-and-trajectory prediction framework, with staged training over Text-to-Action pretraining, Continued Pretraining, Supervised Fine-Tuning, and Reinforcement Learning. Reported results for Qwen-VLA-Instruct include 97.9% on LIBERO, 73.7% on Simpler-WidowX, 86.1%/87.2% on RoboTwin-Easy/Hard, 69.0% OSR on R2R, 59.6% SR on RxR, 76.9% average OOD success in real-world ALOHA experiments, and 26.6% zero-shot success on DOMINO.

Qwen-RobotManip extends this embodied trajectory further by building a manipulation foundation model on Qwen-VL with a unified alignment framework across representation, motion, and behavior (Yuan et al., 16 Jun 2026). The report states that it constructs a ~38,100-hour pretraining corpus using open-source robot datasets, egocentric human videos, and a human-to-robot synthesis pipeline spanning 15 platforms. It reports substantial gains over prior systems on OOD benchmarks, including 91.4% on LIBERO-Plus (avg), 69.4% on RoboTwin-C2R (Hard), 35.9% on RoboCasa365 (Total), 45.6% on EBench (avg), 72.2% on RoboTwin-IF, and 23.9% on RoboTwin-XE (avg XSembod). In this branch of the literature, the Qwen multimodal stack functions as a perception-and-reasoning substrate for action models rather than only as an image-text assistant.

6. Privacy, comparative position, and recurrent misconceptions

One line of work studies privacy risks directly. "DP$^2$-VL: Private Photo Dataset Protection by Data Poisoning for Vision-LLMs" introduces identity-affiliation learning as a threat model in which an attacker fine-tunes a VLM using only a few private photos of a target individual and associated descriptions (Miao et al., 25 Mar 2026). The paper states that mainstream VLMs including Qwen-VL can recognize facial identities and infer identity-affiliation relationships by fine-tuning on small-scale private photographic dataset, and even on synthetically generated datasets. Without protection, ASR$_{id}$ and ASR$_{aff}$ are described as very high, with 90%+ in several scenarios. Under DP$^2$-VL-protected images and LoRA fine-tuning at 100% protection ratio, the reported Qwen-VL values include ASR$_{id}$=3.12% for Brief Prompts and ASR$_{id}$=8.29% for Complex Prompts in the identity scenario. The paper thus places privacy leakage among the nontrivial liabilities of Qwen-class multimodal models.

Comparative reports also show that Qwen VL Plus is strong but not uncontested. Kimi-VL reports that Kimi-VL outperforms Qwen2.5-VL-7B on 19/24 benchmarks tested, with advantages in long-context multimodal QA, agent tasks such as OSWorld and WindowsAgentArena, and several OCR and video benchmarks (Team et al., 10 Apr 2025). MiMo-VL-7B-RL reports that it outperforms Qwen2.5-VL-7B on 35 out of 40 evaluated tasks, including 56.1 versus 37.5 on OSWorld-G (Team et al., 4 Jun 2025). AndesVL reports that its mobile-side models outperform Qwen2-VL and Qwen2.5-VL variants of similar size across text-rich image understanding, general VQA, hallucination mitigation, multi-image understanding, GUI-related tasks, and multilingual evaluation, while also adding explicit mobile deployment mechanisms such as 1+N LoRA, QALFT, OKV cache eviction, and speculative decoding (Jin et al., 13 Oct 2025).

Several common misconceptions are directly challenged by these papers. The Change VQA study shows that larger model size does not guarantee improved performance and that native multimodal models are more effective than structured vision-language pipelines under matched LoRA adaptation (Bazi et al., 20 Apr 2026). GazeQwen argues that learning where to inject gaze is more effective than scaling model size or engineering better prompts (Pham et al., 26 Mar 2026). MMSearch-Plus shows that access to tools and search does not automatically produce stronger multimodal agents; for Qwen-2.5-VL-72B-Instruct, additional search rounds reduced accuracy rather than increasing it (Tao et al., 29 Aug 2025). Taken together, these results indicate that the main determinants of performance are increasingly multimodal backbone integration, alignment strategy, tool-use policy, and task-specific adaptation protocol, not parameter count alone.