InternVL2: Open-Source Vision-Language Backbone

• InternVL2 is an open-source family of multi-scale vision-language models featuring variants from 1B to 76B parameters, adaptable to diverse modalities.
• It integrates vision and language through techniques like tile-based processing, late fusion, and adaptive token compression to enhance multimodal reasoning.
• Extensible for applications such as OCR augmentation, long-context reasoning, and video/audio integration, it serves as both a robust baseline and customizable backbone.

Searching arXiv for InternVL2 and related papers to ground the article in recent literature. arXiv search query: InternVL2 InternVL2 is an open-source family of vision-LLMs that appears in recent literature as a VLM, MLLM, or LVLM backbone for image, document, and video reasoning. Across comparative and systems papers, evaluated variants include 1B, 2B, 4B, 8B, 26B, 40B, and 76B, and the family is repeatedly used both as a state-of-the-art baseline and as a substrate for plug-in OCR, preference optimization, token compression, and long-context extensions (Adil et al., 12 Apr 2025, Schmalfuss et al., 3 Jun 2025, Nacson et al., 2024, Wang et al., 2024).

1. Family definition and nomenclature

InternVL2 is treated in the surveyed literature as a multi-scale open-source model family. Construction-hazard experiments explicitly evaluate 1B, 2B, 4B, and 8B variants and emphasize their suitability for local or in-house deployment, privacy-sensitive settings, and lower operational cost than commercial APIs (Adil et al., 12 Apr 2025). Prompt-robustness evaluation extends the family comparison through 40B, and long-video and long-document studies further use 76B variants, including InternVL2-Llama3-76B in psychometric spatial evaluation (Schmalfuss et al., 3 Jun 2025, Huang et al., 24 Apr 2025, Xu et al., 17 Feb 2025).

A recurrent source of ambiguity is naming. In long video-audio comparison, one paper describes the baseline interchangeably as InternVL2 and InternVideo2, while the main quantitative table reports “InternVideo2 (Base)” (Lin et al., 10 Jan 2026). Within that framing, the model serves as a strong late-fusion video-audio foundation baseline rather than a chunked long-video reasoning system. This suggests that “InternVL2” can denote either the broader VLM family or, in some secondary discussions, a closely related baseline used in video-audio evaluation.

The literature therefore does not present InternVL2 as a single fixed artifact. Instead, it appears as a scalable backbone family whose identity is partly defined by context: document reading, prompt robustness, long-video inference, hallucination mitigation, or domain-specific adaptation.

2. Operating characteristics and architectural interpretations

Several studies characterize InternVL2 through its native handling of visual tokens. In document understanding, it is described as using a tile-based visual processing strategy; under tight token budgets, that strategy can become brittle because fine-grained textual detail and global layout cues may be missed (Nacson et al., 2024). The same study evaluates InternVL2 at 256 visual tokens, corresponding to a single-tile configuration, and at 1280 visual tokens, corresponding to a four-tile configuration, both under a 448 × 448 constrained setting (Nacson et al., 2024).

In video-audio comparison, the model is described as a single late-fusion foundation model that integrates video and audio/text information through late fusion and aims for joint grounding of visual and auditory content in text (Lin et al., 10 Jan 2026). By contrast, later systems such as QMAVIS treat this design as optimized for standard video understanding settings rather than explicit chunked reasoning over very long videos.

Explanations for InternVL2’s empirical behavior are partly architectural and partly data-centric. A distraction-robustness study argues that InternVL2 likely benefits from the InternViT vision encoder, the InternLM2 LLM, and a diverse QA-heavy training mix including COCO, VQAv2, OKVQA, Visual Dialog, ScienceQA, DocVQA, OCR-VQA, TextVQA, and ChartQA (Liu et al., 13 Feb 2025). A context-robustness study proposes that dynamic resolution mechanisms, hierarchical attention layers, and better selective encoding help larger InternVL2 variants remain stable under patch-vs-full-image changes, while still noting that such explanations are diagnostic rather than causal proofs (Patel et al., 28 Sep 2025).

3. Efficiency and document-centric extensions

A substantial portion of the recent InternVL2 literature treats the family as a frozen or lightly modified backbone to which specialized modules can be attached. Three representative lines of work are OCR augmentation, learned visual token compression, and training-free layer-wise token reduction.

Method	Integration with InternVL2	Representative reported effect
DocVLM	Frozen InternVL2 + OCR encoder + 64 learned queries	DocVQA 56.0 to 86.6 at 256 visual tokens
FCoT-VL	Distilled compressed student initialized from InternVL2	InternVL2-8B at 50% compression: DocVQA 91.88, ChartQA 85.52
G-Search / P-Sigmoid	Training-free layer-wise Sort-Reduce inside the LLM backbone	InternVL2-8B TFLOPs 24.10 to 12.24, avg acc. 70.83 to 70.10

DocVLM adds an OCR-side modality while preserving the original InternVL2 weights. It uses the encoder component of DocFormerV2, omits the visual branch to avoid redundancy, compresses OCR text and layout into $M=64$ learned queries, projects them to the VLM hidden size, and concatenates them with visual tokens before the frozen LLM. On InternVL2, this raises DocVQA from 56.0 to 86.6 at 256 visual tokens and from 85.7 to 91.0 at 1280 visual tokens, while also improving TextVQA, ST-VQA, InfoVQA, TextCaps, MP-DocVQA, and DUDE (Nacson et al., 2024).

FCoT-VL addresses high-resolution text-oriented reasoning by using InternVL2 as both teacher and initialization source. In the re-alignment stage, the student ViT and LLM remain frozen while only the student projector and compression module are updated; a later post-train stage makes all parameters learnable. Reported results show that InternVL2-8B with 50% compression still reaches 91.88 on DocVQA, 85.52 on ChartQA, 78.95 on TextVQA, 83.9 on OCRBench, and 63.3 on MathVista, with 103.43% average relative performance versus the original baseline (Li et al., 22 Feb 2025).

A complementary line of work reduces visual token cost without retraining. G-Search inserts a prompt-aware Sort-Reduce module before LLM layers and greedily finds layer-wise keeping rates; for InternVL2-8B it reduces TFLOPs from 24.10 to 12.24 while average accuracy changes from 70.83 to 70.10. Under fixed budgets, P-Sigmoid further improves average accuracy over FastV by +7.69 on InternVL2-8B at similar budgets (Zhao et al., 2024).

4. Long-context, long-video, and multi-page reasoning

Long-context modification is a central theme in the InternVL2 ecosystem. Variable Visual Position Encoding (V2PE) starts from InternVL2-2B and changes only the positional indexing scheme for visual tokens. Text tokens retain increment 1, while visual tokens advance by a smaller increment $\delta < 1$:

$$p_i = p_{i-1} + \begin{cases} 1, & \text{if } x_i \text{ is a textual token}, \ \delta, & \text{if } x_i \text{ is a visual token}. \end{cases}$$

During training, $\delta$ is sampled from $$\{1,\frac{1}{2},\frac{1}{4},\frac{1}{8},\frac{1}{16},\frac{1}{32},\frac{1}{64},\frac{1}{128},\frac{1}{256}\}$$. With this modification and long-context fine-tuning, Long-VQA average rises from 28.4 for base InternVL2-2B to 51.3 at $\delta=1/16$, and MM-NIAH-style performance at 256K tokens improves from 1.5 to 56.9 with $\delta=1/256$. After 256K-token training, the model reaches 95.1 at 512K tokens and 64.5 at 1M tokens (Ge et al., 2024).

FRAG addresses long inputs without long-context retraining by independently scoring sampled frames or pages and selecting Top-$K$ items. Applied zero-shot to InternVL2, it improves long-video performance consistently. For InternVL2-76B, LongVideoBench changes from 59.5 to 61.5, MLVU from 63.3 to 69.2, Video-MME from 62.3 to 66.0, and EgoSchema from 63.1 to 63.8. In long-document benchmarks, FRAG with InternVL2-76B reports SlideVQA EM/F1 of 66.0/75.6, MP-DocVQA ANLS of 88.3, and MMLongBench-Doc Acc/F1 of 37.9/34.8 (Huang et al., 24 Apr 2025).

LVC pushes InternVL2 further toward long-video modeling by sampling 64 frames, compressing them into pseudo-image frames with Query-Attention Video Compression, and training only the alignment layer. InternVL2-40B-LVC reaches 68.2 on MLVU and 65.9 on Video-MME, relative improvements of 14.6% and 7.7% over the corresponding baseline (Wang et al., 9 Apr 2025).

The importance of such augmentations is underscored by QMAVIS. In that comparison, InternVideo2 (Base), discussed in the same paper’s framing as InternVL2/InternVideo2, scores 41.9 on VideoMME (w/ subs), 52.16 on PerceptionTest, and 55.0 on EgoSchema, while QMAVIS reaches 66.46, 57.72, and 65.0. Qualitative comparison in the same work reports that the baseline provides highly detailed character-level descriptions but fails to identify multiple scenes and the overall storyline in long videos (Lin et al., 10 Jan 2026).

5. Robustness profile and failure modes

InternVL2’s strongest comparative reputation in the surveyed literature is prompt robustness. PARC evaluates 22 VLMs under 11 prompt variations and ranks InternVL2 as the most robust family overall, with the 40B model best. PARC defines reliability as

$$\mathit{rel} = (2 \cdot \mathit{acc} - 1) \cdot \mathit{cert},$$

and reports calibrated reliability AVG of 0.09, 0.17, 0.26, 0.32, 0.38, and 0.40 for InternVL2 1B, 2B, 4B, 8B, 26B, and 40B respectively; calibrated consistency AVG rises in parallel from 0.31 to 0.71. The same study emphasizes that InternVL2 is not immune to perturbation: semantic changes remain much harder than reformulations, but the family is less destabilized than most competitors (Schmalfuss et al., 3 Jun 2025).

Robustness to distraction shows a similar pattern. On I-ScienceQA, which injects visual and textual distractions into ScienceQA, InternVL2-8B records distracted accuracies of 94.45 for Add Image, 94.23 for Insert Image, 93.60 for Add Hint, and 95.90 for Insert Hint, corresponding to degradations of -1.00, -2.67, -1.20, and -1.70. InternVL2-26B also remains strong, and both outperform GPT-4o on all four distracted conditions. At the same time, the authors explicitly warn that ScienceQA appears in InternVL2’s training mix, so some of the apparent robustness may reflect contamination rather than purely out-of-distribution generalization (Liu et al., 13 Feb 2025).

Patch Context Robustness Index (PCRI) extends the robustness question from prompt variation to visual context granularity. InternVL2-26B is singled out as one of the few models with near-zero average PCRI across tasks, indicating relatively stable performance between localized patches and full-image input. Yet the same study also reports strongly negative captioning PCRI for InternVL2-26B, with -0.4921 at $2 \times 2$ and -0.7547 at $\delta < 1$0, showing that robustness is task-dependent rather than absolute (Patel et al., 28 Sep 2025).

The clearest weakness emerges in psychometric spatial evaluation. Across nine classic tests mapped to five Basic Spatial Abilities, InternVL2 receives an overall manufacturer-level score of 19.6, below the VLM average of 24.95 and far below the human average of 68.38. The paper further states that Intern-VL2-76B and Intern-VL2-8B completely failed in spatial orientation. This result is used to argue that general multimodal strength does not imply robust spatial intelligence, especially in mental rotation, spatial orientation, and dynamic 3D transformation (Xu et al., 17 Feb 2025).

6. Alignment, reasoning enhancement, and applied deployments

InternVL2 is also a frequent target for post-training alignment. Mixed Preference Optimization (MPO) is designed to improve multimodal Chain-of-Thought reasoning by combining DPO, BCO, and SFT on the MMPR preference dataset. The baseline motivation is explicit: in an ablation, InternVL2-8B scores 58.3 on MathVista with direct answers but only 56.8 with CoT. After MPO, InternVL2-8B-MPO reaches 79.2 on M3CoT, 67.0 on MathVista, 25.7 on MathVision, 56.2 on MMVet, 76.7 on LLaVA-Bench, and 88.1 on POPE, making it approximately comparable to the 10× larger InternVL2-76B on MathVista and stronger than that baseline on M3CoT and MathVision (Wang et al., 2024).

Hallucination-focused alignment yields a different improvement profile. SEED identifies low-confidence internal knowledge, isolates hallucination-prone logits through image perturbation, purifies them by confidence-conditioned subtraction, and distills the purified distribution back into the model through a mode-seeking objective and a Hallucination Elimination Adapter. For InternVL2-8B, POPE-Random F1 improves from 88.11 to 89.92 and MM-Vet from 54.2 to 59.4; for InternVL2-76B, MM-Vet rises from 65.7 to 69.2. The paper emphasizes that these gains are obtained without extra inference passes, unlike Visual Contrast Decoding (Li et al., 7 Jul 2025).

Task-specific studies use InternVL2 as a large multimodal classifier or decision-support backbone. In multimodal hate detection, InternVL2-8B is evaluated with simple versus category prompts, binary versus scaled labels, and prompting-only versus LoRA-based fine-tuning. The best binary F1 is 68.30 with multimodal fine-tuning, category prompt, and binary labels; the best scaled F1 is 64.74 with multimodal fine-tuning, category prompt, and scaled labels (Singh et al., 15 Aug 2025). In construction safety, InternVL2-8B is the strongest family member tested, with hazard-detection Cosine Similarity 0.501, BERTScore F1 0.860, and average inference time 5.30 seconds per image on an NVIDIA A6000 GPU, but it still trails GPT-4o and Gemini 1.5 Pro in overall hazard-identification quality (Adil et al., 12 Apr 2025). In wheat breeding, an InternVL2-8B-based WBLM augmented with SFT, RAG, and RLHF achieves the leading open-source results in that study, including wheat yield prediction with $\delta < 1$1 and RMSE $\delta < 1$2 kg/ha, plus a stability score of 0.811 (Yang et al., 2024).

Collectively, these studies portray InternVL2 as a strong and unusually extensible open-source VLM family: comparatively prompt-robust and adaptable across many downstream regimes, yet still limited in psychometric spatial reasoning, vulnerable to some forms of distraction and contextual brittleness, and often substantially improved by targeted modules or post-training rather than by scale alone.