MagicVL-2B: Mobile Vision-Language Model

• MagicVL-2B is a compact vision-language model designed for on-device deployment, featuring a lightweight 93M visual encoder and tailored for smartphone constraints.
• It employs a redesigned dynamic resolution mechanism and a four-stage multimodal curriculum learning strategy to optimize token efficiency and maintain performance.
• Benchmark results show significant reductions in latency and power consumption while achieving competitive accuracy on tasks such as TextVQA, Document QA, and real-world evaluations.

MagicVL-2B is a compact vision-LLM (VLM) designed specifically for on-device deployment on flagship smartphones. It is presented as an enhanced variant built on the InternVL2.5 framework, combining a lightweight visual encoder with fewer than 100M parameters, a redesigned dynamic resolution mechanism, and a four-stage multimodal curriculum learning strategy intended to recover capability that is often lost when the visual side is made small. The deployment target is explicit: smartphone inference on Snapdragon 8 Elite hardware, with the broader aim of enabling multimodal understanding and autoregressive generation without cloud dependence, while reducing latency, storage pressure, and on-device power consumption (Liu et al., 3 Aug 2025).

1. Mobile-first problem setting and design objectives

MagicVL-2B is motivated by a deployment setting in which memory is limited, mobile processors are compute-constrained, and standard VLM visual encoders are often too large and too costly for phone hardware. The paper frames smartphones as the most ubiquitous and accessible computing platforms, and treats lower latency, better privacy, and broader accessibility as primary reasons to run VLMs directly on device. The motivating application space includes real-time translation, augmented reality, smart assistants, and visual question answering (Liu et al., 3 Aug 2025).

The model addresses three bottlenecks in that setting. First, it reduces visual-side cost through a lightweight encoder. Second, it redesigns dynamic resolution processing so that image tokens are generated adaptively without excessive modification of image dimensions. Third, it introduces curriculum learning so that a compact encoder can still align effectively with a relatively strong LLM. The paper’s central claim is therefore not simply that MagicVL-2B is small, but that it is co-designed for actual smartphone constraints in compute, token budget, and power.

This positioning distinguishes MagicVL-2B from compact multimodal models whose parameter count is reduced without equivalent attention to mobile inference behavior. In the paper’s own framing, mobile-friendliness depends jointly on model scale, visual token generation, and the training regime used to preserve multimodal competence.

2. Architecture and model components

MagicVL-2B is described as an enhanced variant built on the InternVL2.5 framework. Its learnable architecture has three principal components: a visual encoder, a two-layer MLP projector, and a LLM backbone. The model can be instantiated with either Qwen2.5-1.5B or Qwen3-1.7B; unless otherwise specified, “MagicVL-2B” refers to the Qwen3-1.7B version (Liu et al., 3 Aug 2025).

The visual encoder is SigLIP2-Base-384/16 with 93M parameters. The paper reports that three encoder families were evaluated—ViT, SigLIP, and SigLIP2—and that SigLIP2 delivered the best performance among these options. The chosen configuration uses base scale, input resolution $384 \times 384$, and patch size 16. The stated rationale is threefold: the base variant keeps the encoder efficient at about 93M parameters, the 384 resolution improves global visual information, and patch size 16 preserves finer-grained details. This is a deliberate departure from heavier CLIP-style or InternViT-style visual backbones often seen in prior open VLMs.

The multimodal connector is a two-layer MLP projector that maps image tokens into the LLM token space. The paper does not provide a projection equation, but the architectural role is conventional: visual tokens are encoded by the ViT-like tower, transformed into the language embedding space, and then consumed by the autoregressive LLM for multimodal generation.

The model’s token budgeting is also explicit. In training, the packed batch uses a maximum token length of 16,384, up to 48 images, and a maximum of 24 dynamic patches. These limits matter because visual sequence length affects not only the encoder but also LLM attention cost, memory footprint, and downstream latency.

3. Redesigned dynamic high-resolution mechanism

A major systems contribution of MagicVL-2B is its redesigned dynamic high-resolution scheme. The paper argues that prior dynamic resolution methods in InternVL-style systems often resize images so height and width become integer multiples of the pretraining resolution, which can distort aspect ratios, especially for unusual mobile inputs such as long screenshots, and can also generate redundant tokens. MagicVL-2B instead resizes each dimension to the nearest integer multiple of the pixel size corresponding to one visual token, so resizing is performed at token granularity rather than at the granularity of the full pretraining resolution (Liu et al., 3 Aug 2025).

The paper’s resizing equation is typeset imperfectly in the provided text, but its intended semantics are described clearly. The patch size is 16, and token compression uses pixel unshuffle with compression ratio $R_{\text{psf}} = 2$. The resizing rule rounds each image dimension to the nearest multiple of a token-sized spatial unit. The stated consequence is that the method minimizes distortion, preserves original content more faithfully, and avoids generating unnecessary visual tokens.

Because the visual encoder expects fixed-size input at $384 \times 384$, the method adds a padding-and-masking stage. Images are padded with zeros up to $384 \times 384$ when needed, an attention mask is applied, tokens from padded regions are discarded, and only tokens corresponding to original image content are fed to the LLM. This prevents padded regions from acting as spurious content while retaining compatibility with the fixed-resolution encoder.

The paper reports a direct ablation on TextVQA. Under InternVL2.5-2B dynamic resolution, token consumption is 0.82M and TextVQA is 74.3. Under MagicVL-2B dynamic high resolution, token consumption is 0.51M and TextVQA is 74.5. The paper describes this as about 37.8% fewer tokens with slightly better accuracy. In practical terms, that reduction lowers visual encoder work, shortens multimodal sequences processed by the LLM, and plausibly contributes to both latency and power gains.

4. Data curation and curriculum learning

The training corpus is approximately 150 million image-text pairs curated from open-source datasets. The paper emphasizes high data quality, broad visual diversity, and bilingual capability; because open-source Chinese image-text data are limited, some English data are translated into Chinese using large LLMs. Data filtering proceeds in three stages: heuristic rule-based filtering, rule-based duplication detection, and LLM prompt-based filtering for logical coherence and likely hallucinations (Liu et al., 3 Aug 2025).

Datasets are categorized into reasoning, GUI, OCR, text-only, chart, caption, visual question answering, and grounding. If a dataset has explicit task labels, those labels are used directly. If not, manual inspection is used first; if the dataset is mixed, a large VLM is used to classify samples into task-specific subsets. This categorization is not merely administrative: it underpins the paper’s complexity-aware curriculum.

The curriculum is formalized over datasets of the form

$$\mathcal{D} = \{(v_i, p_i, x_i) \mid i = 1, \ldots, n\},$$

where $v_i$ is the image, $p_i$ the prompt, and $x_i$ the response. Textual information complexity is measured using average response length,

$L = \frac{1}{n}\sum_{i=1}^n \text{len}(x_i),$

type-token ratio,

$T = \frac{1}{n}\sum_{i=1}^n \text{TTR}(p_i+x_i),$

and perplexity,

$R_{\text{psf}} = 2$0

Perplexity is computed with Qwen2.5-1.5B. Visual information complexity is based on image entropy, text density, and object density. Cross-modal task complexity follows an NVILA-style loss-gap comparison across different model sizes, using Qwen2-VL 2B, 7B, and 72B. The overall complexity score is a weighted combination of text, image, and task complexity, with task-dependent weights provided in supplementary material.

The curriculum then unfolds in four stages. Stage 1, Foundational Modality Alignment, freezes both the visual encoder and the LLM and updates only the MLP projector on 10M low-complexity image-caption pairs. Stage 2, Enhanced Visual Representation, trains the visual encoder and MLP projector while keeping the LLM frozen, using 23M high-complexity image-caption pairs. Stage 3, Generalized Multi-Modal Ability, unfreezes all components and trains on 54M low-complexity multimodal instruction-following samples. Stage 4, Advanced Multi-Modal Ability, jointly trains all components on 66M high-complexity data across all tasks.

The paper argues that this progression incrementally increases both information density and task difficulty. Its stated purpose is to stabilize convergence, establish basic alignment first, improve visual representation next, and only then expose the full model to harder instruction following and reasoning. This training strategy is a central part of the model’s identity: the compact visual encoder is not treated as a mere efficiency swap, but as a component that requires a specialized training schedule.

5. Optimization, training configuration, and on-device deployment

Optimization uses AdamW with cosine decay. Training is performed on 128 NVIDIA A800 80G GPUs. The packed batch retains the previously stated limits of maximum token length 16,384, up to 48 images, and maximum 24 dynamic patches. Stage-specific hyperparameters are explicit: Stage 1 uses learning rate $R_{\text{psf}} = 2$1, warmup 100 steps, and 65k training steps; Stage 2 uses $R_{\text{psf}} = 2$2, warmup 100 steps, and 90k steps; Stage 3 uses $R_{\text{psf}} = 2$3, warmup ratio 0.03, and 140k steps; Stage 4 uses $R_{\text{psf}} = 2$4, warmup ratio 0.03, and 250k steps. The paper does not specify RLHF, DPO, quantization-aware training, or other post-training alignment stages (Liu et al., 3 Aug 2025).

The deployment evaluation is performed on Snapdragon 8 Elite hardware in a head-to-head comparison with InternVL2.5-2B. The paper does not specify the inference framework, quantization format, runtime library, memory footprint, compression method, or exact phone model, which limits deployment reproducibility from the main text alone.

Metric	InternVL2.5-2B	MagicVL-2B
Model loading	1.04 s	1.01 s
ViT latency	0.90 s	0.09 s
LLM latency	2.0 s	1.7 s
Throughput	14.3 token/s	23.9 token/s

The most striking result is the 10× reduction in ViT latency, from 0.90 s to 0.09 s. LLM latency also decreases, from 2.0 s to 1.7 s, which is consistent with a shorter visual token sequence. Throughput rises from 14.3 to 23.9 token/s, roughly 1.67× higher, while model loading time remains essentially unchanged.

The paper also reports a headline reduction of on-device power consumption by 41.1%. However, the detailed comparison table, measurement protocol, instrumentation, and precise baseline conditions for that figure are not described in the visible main text. The most defensible reading is therefore that the model is reported to reduce inference power substantially, at least relative to InternVL2.5-2B under on-device conditions, while the exact measurement methodology is not specified in the provided text.

6. Benchmark results, comparative context, and limitations

MagicVL-2B is evaluated on HallusionBench, CRPE, MMBench_V11_en, RealworldQA, MME_Realworld, MMStar, DocVQA, OCRBench, AI2D, and SEED-2 Plus. The paper reports results for both backbone variants (Liu et al., 3 Aug 2025).

Benchmark	Qwen2.5-1.5B	Qwen3-1.7B
HallusionBench	47.7	50.8
CRPE	70.9	70.3
MMBench	71.8	73.7
RealworldQA	61.4	63.5
MME_Realworld	49.8	49.1
MMStar	52.7	57.9
DocVQA	87.7	89.0
OCRBench	775	828
AI2D	76.7	77.4
SEED-2 Plus	61.0	61.6

The Qwen3-1.7B variant is generally stronger overall. Among models at or below 2B parameters, the paper states that MagicVL-2B achieves the top score on HallusionBench, MMBench, RealworldQA, MMStar, OCRBench, and AI2D. Relative to InternVL2.5-2B, it improves on HallusionBench, MMBench, RealworldQA, MME_Realworld, MMStar, DocVQA, OCRBench, AI2D, and SEED, while roughly matching CRPE. The paper also highlights that some of these results exceed those of larger models, although it does not claim universal superiority over larger systems.

The curriculum ablation is consistent with the model’s training thesis. Without data categorization and progressive training, HallusionBench, CRPE, MME_Realworld, and DocVQA are 49.5, 69.2, 48.2, and 87.5. With data categorization but no progressive training, they become 50.3, 70.0, 48.4, and 88.1. With the full curriculum, they reach 50.8, 70.3, 49.1, and 89.0. The pattern supports the claim that data categorization already helps, and progressive curriculum adds further gains.

The limitations are also clear. MagicVL-2B does not uniformly beat larger 7B–8B class systems; it remains below stronger models on tasks including CRPE, some MMBench comparisons, DocVQA, AI2D, and SEED. The paper does not provide extensive qualitative failure analysis, a detailed safety discussion, exact mobile system-stack details, exact power measurement protocol, or memory-footprint numbers. These omissions matter for reproducibility and for operational assessment in production mobile systems.

A useful indirect comparison point is AuroraEdge-V-2B, another compact 2B-class multimodal model, but one targeted at edge and industrial deployment rather than smartphones. AuroraEdge-V-2B uses a LLaVA-derived compression-fusion design that compresses 256 visual tokens to 64 before decoding and reports 40 ms latency on an RTX 3090, whereas MagicVL-2B emphasizes a lightweight 93M visual encoder, token-level dynamic resizing, and smartphone deployment on Snapdragon 8 Elite (Chen, 23 Jan 2026). This suggests that contemporary compact VLM design is splitting along at least two optimization axes: mobile-first reduction of visual-encoder and token-generation cost, and edge-first reduction of decode-time visual-token burden.

Taken together, MagicVL-2B is best understood as a mobile-first 2B-scale VLM whose significance lies in the joint design of architecture, tokenization, and curriculum. Its contribution is not merely to shrink a server-side multimodal model, but to show that a smartphone-oriented VLM can pair a 93M SigLIP2-Base visual encoder, a two-layer MLP projector, a compact Qwen-family LLM, and a complexity-aware four-stage curriculum to achieve strong benchmark coverage with materially improved on-device efficiency (Liu et al., 3 Aug 2025).