[whatmodel.sh](../../README.md) / InternVL2-4B # InternVL2-4B Shanghai AI Laboratory (OpenGVLab) · released 2024-07 · 4.15B params · mit **score 42.5** ±1.9 · #526 of 732 · from 19 benchmark results across 4 sources Score within the [LMArena (text)](https://whatmodel.sh/?set=lmarena-text&score=set) comparison set only. ## What it's good at _{● marks InternVL2-4B among every model on that task or test, worst to best.}

task                vs the others          rank  counts  tests
math                ▁▁▃▅●▆▄█▄▂▁▁   #261 of 431*     14%      1
vision & documents  ▁●▂▂▁▃▄▂▆▇█▄    #148 of 153     84%     13

math · #261 of 431 · 14% of its score

Beats [GPT-4o (2024-11-20)](gpt-4o-2024-11-20.md), [GPT-4 Turbo (2024-04-09)](gpt-4-turbo-2024-04-09.md), [Llama 3.1 405B Instruct](llama-3-1-405b.md) here, though they rank higher overall. Trails [DeepSeek R1 Distill Qwen 1.5B](deepseek-r1-distill-qwen-1-5b.md), [Gemma 7B (base)](gemma-7b-base.md), [CodeGemma 7B (v1.0)](codegemma-7b.md) here, though they rank lower overall.

test                  result  vs the others in its run  difficulty        counts  why
MathVista (testmini)   53.7%  ▁··▁▄▄●▇▄▅█▄ beats 12/48  medium · avg 65%     14%  at its level · small run · several runs

vision & documents · #148 of 153 · 84% of its score

Beats [Gemma 3 4B](gemma-3-4b.md), [Phi-3-vision-128k-instruct](phi-3-vision-128k-instruct.md) here, though they rank higher overall. Trails [Qwen3.5 0.8B](qwen3-5-0-8b.md), [North Micro Vision Instruct (2.4B)](north-micro-vision-instruct.md), [Molmo 7B-D (0924)](molmo-7b-d.md) here, though they rank lower overall.

test                                result  vs the others in its run  difficulty           counts  why
Arena vision (style controlled)      958.4  ▁▁●▂▂▂▃▄▅▅█▇ beats 2/152  avg 1061                24%  blunt
MMMU (val)                           44.2%  ▁····▁▁●█▆▆▅ beats 6/53   medium · avg 51%        19%  several runs
MathVista (testmini)                 53.7%  ▁··▁▄▄●▇▄▅█▄ beats 12/48  medium · avg 65%        14%  at its level · small run · several runs
ChartQA                              78.8%  ▁··▁▁▁▁▁▂●▅█ beats 17/44  easy · avg 82%         4.6%  blunt · small run · several runs
DocVQA (maker)                       89.2%  ▁··▃●▁▁▇█▇▄▅ beats 3/28   very easy · avg 93%    3.8%  blunt · small run
PhysBench physical-world understa…   39.7%  ▅▃●··▅▃····▃ beats 4/9    hard · avg 42%         3.7%  blunt
HallusionBench (maker)               41.9%  ●▁·▂▂▂▃▅▃█▅▅ beats 1/33   medium · avg 55%       3.6%  blunt · at its level · small run
TextVQA                              66.2%  ▁▁▁▄▅▁●·▃▃▃█ beats 11/24  easy · avg 75%         3.2%  blunt · at its level · several runs
RealWorldQA (maker)                  60.7%  ▁▃●▇▃▇▇▅█▁▅▅ beats 3/39   easy · avg 71%         2.9%  blunt · at its level · small run
AI2D test                            77.3%  ▁▁▅▇●▇·▄·█·▃ beats 11/28  easy · avg 84%         2.6%  blunt · small run · several runs
OCRBench (maker)                     78.8%  ▁·▃▁●▅▂█▅█▁▃ beats 6/37   easy · avg 83%         2.4%  blunt · small run
CharXiv (reasoning questions) (ma…   24.5%  ●▅▃▅▅▆▆▅▆█▆▂ beats 1/101  hard · avg 34%            —  not counted yet: no other model in the …
CharXiv descriptive questions (ma…   48.0%  ●▄▄▅▇▄██▇▂▁▂ beats 0/59   easy · avg 69%            —  not counted yet: no other model in the …

other tests (3)

test                                result  vs the others in its run  difficulty        counts  why
Video-MME (with subtitles) (maker)   53.4%  ●▁▁▁▁·▁▃▁▁▁█ beats 0/16   medium · avg 66%    6.5%  at its level · small run
Video-MME (without subtitles) (ma…   51.4%  ●▃▁▃▁·█▄▄▅·▇ beats 0/27   medium · avg 63%    5.0%  blunt · at its level · small run
MVBench (maker)                      63.7%  ▁··▃●▁▃▄▄▄█▄ beats 3/26   medium · avg 67%    2.6%  blunt · at its level · small run

## How the score was worked out Each result implies a score on its own, after adjusting for how hard that test is and who ran it. The model's score is their weighted average, with a small pull toward the middle for thinly tested models. Each test's weight is its *counts* column above; here it is added up by who ran the tests. [Method](../method.md)

source                                    results  share of the score
OpenGVLab InternVL2 model cards (submit…       12   52% ▇▇▇▇▇▇
LMArena leaderboard dataset                     1   24% ▇▇▇
Microsoft Phi-3.5-vision-instruct model…        5   19% ▇▇
PhysBench: Benchmarking and Enhancing V…        1    4% ▇
prior: pull toward the middle                        2% ▇
weighted average                                   = 42.5

_{Licences: OpenGVLab InternVL2 model cards ([MIT](https://huggingface.co/OpenGVLab/InternVL2-26B/blob/46f37972e05604cc44a59f72334df220fa197a7b/README.md)) · LMArena leaderboard dataset ([CC BY 4.0](https://creativecommons.org/licenses/by/4.0/)) · Microsoft Phi-3.5-vision-instruct model card: Benchmarks table ([MIT](https://huggingface.co/microsoft/Phi-3.5-vision-instruct/blob/12b77fb40b63a2c73c68243d3f767aab688a1b2a/LICENSE)) · PhysBench: Benchmarking and Enhancing Vision-Language Models for Physical World Understanding ([CC-BY-4.0](https://creativecommons.org/licenses/by/4.0/))}

Not counted (2)

source     test                           score  why
OpenGVLab  CharXiv (reasoning questions)  24.5%  too few models have it yet
OpenGVLab  CharXiv descriptive questions  48.0%  too few models have it yet

## How to run it ### Other ways to run it 1. Run it as published (BF16, needs ~8.29 GB, NVIDIA only, from [github.com/vllm-project](https://github.com/vllm-project/vllm/blob/64cb68384241f046d07ba3797e1b14dbc82a0b2d/docs/models/supported_models.md)) ```sh vllm serve OpenGVLab/InternVL2-4B --trust-remote-code ``` Model repository and documentation: [huggingface.co/OpenGVLab](https://huggingface.co/OpenGVLab/InternVL2-4B) [interactive version](https://whatmodel.sh/model/internvl2-4b) · [JSON](https://whatmodel.sh/api?format=json&all=1)