[whatmodel.sh](../../README.md) / InternVL2-26B # InternVL2-26B Shanghai AI Laboratory (OpenGVLab) · released 2024-07 · 25.5B params · mit **score 47.8** ±1.5 · #412 of 732 · from 17 benchmark results across 5 sources Score within the [LMArena (text)](https://whatmodel.sh/?set=lmarena-text&score=set) comparison set only. ## What it's good at ● marks InternVL2-26B among every model on that task or test, worst to best. 
task                vs the others          rank  counts  tests
math                ▁▁▃▅●▆▄█▄▂▁▁   #270 of 431*    5.8%      1
vision & documents  ▁▁▂●▁▃▄▂▆▇█▄    #132 of 153     90%     14
math · #270 of 431 · 5.8% of its score Beats [GPT-4 Turbo (2024-04-09)](gpt-4-turbo-2024-04-09.md), [Molmo 2 8B](molmo-2-8b.md), [Llama 3.3 70B Instruct](llama-3-3-70b.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) (maker)   59.4%  ▁··▁▄▄●▇▄▅█▄ beats 15/48  medium · avg 65%    5.8%  at its level · small run
vision & documents · #132 of 153 · 90% of its score Beats [Nova Pro](nova-pro.md), [Nova Lite](nova-lite.md) here, though they rank higher overall. Trails [Qwen3.5 2B](qwen3-5-2b.md) here, though they rank lower overall. 
test                                 result  vs the others in its run   difficulty           counts  why
CharXiv descriptive questions         62.4%  ▄█▄●▄·██·▄·█ beats 4/12    easy · avg 69%          28%  at its level
Arena vision (style controlled)      1024.8  ▁▁▂▂●▂▃▄▅▅█▇ beats 16/152  avg 1061                15%  blunt
CharXiv (reasoning questions)         33.4%  ▂▂▂●▃▂·▂▂··▂ beats 7/13    hard · avg 34%          13%  at its level
CapArena-Auto detailed image capt…  13.0000  ▃▃▃·▃█·●▃▃▃▃ beats 8/12    avg -4.498             9.4%  typical
MMMU (val) (maker)                    51.2%  ▁····▁▁▃●▆▆▅ beats 11/53   medium · avg 57%       6.2%  at its level · small run
MathVista (testmini) (maker)          59.4%  ▁··▁▄▄●▇▄▅█▄ beats 15/48   medium · avg 65%       5.8%  at its level · small run
PhysBench physical-world understa…    43.5%  ▅▃█··●▃····▃ beats 7/9     hard · avg 42%         2.6%  blunt
HallusionBench (maker)                50.7%  ▂▁·●▂▂▃▅▃█▅▅ beats 4/33    medium · avg 55%       2.2%  blunt · at its level · small run
DocVQA (maker)                        92.9%  ▁··▃▁▁▁●█▇▄▅ beats 7/28    very easy · avg 93%    2.0%  blunt · maxed out · small run
ChartQA (maker)                       84.9%  ▁··▁▁▁▁▁▂▅●█ beats 24/44   easy · avg 82%         2.0%  blunt · small run
RealWorldQA (maker)                   68.3%  ▁▃▁▇●▇▇▅█▁▅▅ beats 9/39    easy · avg 71%         1.7%  blunt · at its level · small run
OCRBench (maker)                      82.5%  ▁·▃▁▆●▂█▅█▁▃ beats 13/37   easy · avg 83%         1.4%  blunt · small run
AI2D test (maker)                     84.5%  ▁▁▅▇▃●·▄·█·▃ beats 16/28   easy · avg 84%         0.9%  blunt · small run
TextVQA (maker)                       82.3%  ▁▁▁▄▅▁▃·▃●▃█ beats 16/24   easy · avg 77%         0.8%  blunt · small run
other tests (3) 
test                                result  vs the others in its run  difficulty        counts  why
Video-MME (with subtitles) (maker)   57.1%  ▁●▁▁▁·▁▃▁▁▁█ beats 1/16   medium · avg 66%    4.1%  at its level · small run
Video-MME (without subtitles) (ma…   54.8%  ▁●▁▃▁·█▄▄▅·▇ beats 1/27   medium · avg 63%    3.2%  blunt · at its level · small run
MVBench (maker)                      67.5%  ▁··▃▄▁●▄▄▄█▄ beats 7/26   medium · avg 67%    1.6%  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
CharXiv leaderboard                             2   40% ▇▇▇▇▇
OpenGVLab InternVL2 model cards (submit…       12   32% ▇▇▇▇
LMArena leaderboard dataset                     1   15% ▇▇
CapArena: Benchmarking and Analyzing De…        1    9% ▇
PhysBench: Benchmarking and Enhancing V…        1    3% ▇
prior: pull toward the middle                        1% ▇
weighted average                                   = 47.8
Licences: CharXiv leaderboard ([Apache-2.0](https://www.apache.org/licenses/LICENSE-2.0)) · 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/)) · CapArena: Benchmarking and Analyzing Detailed Image Captioning in the LLM Era ([CC-BY-4.0](https://creativecommons.org/licenses/by/4.0/)) · PhysBench: Benchmarking and Enhancing Vision-Language Models for Physical World Understanding ([CC-BY-4.0](https://creativecommons.org/licenses/by/4.0/)) ## How to run it ### Other ways to run it 1. Run it as published (BF16, needs ~51.03 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/InternVL2-26B --trust-remote-code ``` Model repository and documentation: [huggingface.co/OpenGVLab](https://huggingface.co/OpenGVLab/InternVL2-26B) [interactive version](https://whatmodel.sh/model/internvl2-26b) · [JSON](https://whatmodel.sh/api?format=json&all=1)