InternVL2-4B

+ contribute

InternVL2-4B is an open-weights language model from Shanghai AI Laboratory (OpenGVLab), released July 2024. Its published size is 4.15B parameters. Licence: mit.

Catalogue identity notes

Comparison-only. Arena entry 'internvl2-4b' (vision). LMArena's leaderboard table (arena-leaderboard space @6f2eeeb) labels 'internvl2-4b' as 'InternVL2-4b' (OpenGVLab, MIT), linking https://internvl.github.io/blog/2024-07-02-InternVL-2.0/. Maker repo OpenGVLab/InternVL2-4B @6eb21ec: licence tag mit. 4,146,855,936 parameters from the Hugging Face safetensors metadata (dense, so active = total). Released 2024-07: InternVL 2.0 blog post internvl.github.io/blog/2024-07-02-InternVL-2.0/ (linked by LMArena's table).

41.2±1.9#512 of 719 LLMsfrom 19 benchmark results across 4 sources

$ how it compares

The LLMs ranked either side of InternVL2-4B, out of 828 with a score, then more from Shanghai AI Laboratory (OpenGVLab).

#	model	score	where	RAM	released
602	Phi-3-vision-128k-instruct Microsoft	41±2	local	~10GB	2024-05
603	OpenChat 3.5 OpenChat	41±2	local	~6.1GB	2023-10
604	InternVL2-4B Shanghai AI Laboratory (OpenGVLab)	41±2	local	~10GB	2024-07
605	Qwen1.5 14B Chat Alibaba	41±1	local	~11GB	2024-02
606	Code Llama 34B Python Meta	41±1	local	~23GB	2023-08
	more from Shanghai AI Laboratory (OpenGVLab)
479	InternVL2-26B Shanghai AI Laboratory (OpenGVLab)	47±2	local	~55GB	2024-07

see all LLMs · find the ones that fit your machine

$ what it's good at

math▁▂▃▅▆▅▄█▄▁▁▁ #261 of 398 * · 15% of its score · 1 test

Beats Molmo 2 8B, Llama 3.3 70B Instruct, DeepSeek-V2.5 here, though they rank higher overall. Trails DeepSeek R1 Distill Qwen 1.5B, Mixtral 8x7B v0.1 (base) here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
MathVista (testmini) +1	53.7% ✎	▁··▁▄▄▇█▅▆█▄ beats 12/47	medium · avg 66%	▇▇▇▇▇▇ 15%	at its level · small run · several runs

vision & documents▁▃▃▂▁▃▄▂▆▇█▄ #142 of 147 · 84% of its score · 13 tests

Beats Gemma 3 4B, Phi-3-vision-128k-instruct here, though they rank higher overall. Trails Qwen3.5 0.8B, Molmo 7B-D (0924), North Micro Vision Instruct (2.4B) here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
Arena vision (style controlled)	958.3 ✎	▁▁▃▂▂▂▄▅▅▆██ beats 2/145	avg 1068	▇▇▇▇▇▇▇▇▇ 22%	blunt
MMMU (val) +1	44.2% ✎	▁····▁▁▄█▇▆▆ beats 6/52	medium · avg 52%	▇▇▇▇▇▇▇▇ 20%	several runs
MathVista (testmini) +1	53.7% ✎	▁··▁▄▄▇█▅▆█▄ beats 12/47	medium · avg 66%	▇▇▇▇▇▇ 15%	at its level · small run · several runs
ChartQA +1	78.8% ✎	▁··▁▁▁▁▁▂▅▅█ beats 17/44	easy · avg 82%	▇▇ 4.5%	blunt · small run · several runs
DocVQA lab	89.2% ✎	▁··▃▃▁▁▇█▇▄▅ beats 3/28	very easy · avg 93%	▇▇ 3.8%	blunt · small run
PhysBench physical-world understanding (test, Avg)	39.7% ✎	▅▃█··▅▃····▃ beats 4/9	hard · avg 43%	▇ 3.7%	blunt
HallusionBench lab	41.9% ✎	▃▁·▃▃▃▄▅▄█▅▅ beats 1/32	medium · avg 56%	▇ 3.7%	blunt · at its level · small run
TextVQA +1	66.2% ✎	▁▁▁▄▅▁▃·▃▃▃█ beats 11/24	easy · avg 76%	▇ 3.1%	blunt · at its level · several runs

5 more tests

test	result	vs the others in its run	difficulty	counts	why
RealWorldQA lab	60.7% ✎	▂▃▃█▃██▆█▂▆▆ beats 3/38	easy · avg 71%	▇ 2.8%	blunt · at its level · small run
OCRBench lab	78.8% ✎	▁·▃▁▆▅▂█▅█▁▃ beats 6/37	easy · avg 83%	▇ 2.4%	blunt · small run
AI2D test +1	77.3% ✎	▁▁▅▇▃▇·▄·█·▃ beats 11/28	easy · avg 84%	▇ 2.4%	blunt · at its level · small run · several runs
CharXiv (reasoning questions) lab	24.5% ✎	▃▅▃▅▅▆▆▅▆█▆▂ beats 1/100	medium · avg 36%	—	not counted yet: no other model in the same run to compare it with
CharXiv descriptive questions lab	48.0% ✎	▃▄▄▅▇▄█▇▇▂▁▂ beats 0/58	easy · avg 71%	—	not counted yet: no other model in the same run to compare it with

other tests3 not in a task · 15% of its score

test	result	vs the others in its run	difficulty	counts	why
Video-MME (with subtitles) lab	53.4% ✎	▃▁▁▁▁·▁▃▁▁▁█ beats 0/16	medium · avg 68%	▇▇▇ 7.1%	at its level · small run
Video-MME (without subtitles) lab	51.4% ✎	▃▃▁▃▁·█▄▄▅·▇ beats 0/27	medium · avg 64%	▇▇ 5.1%	blunt · at its level · small run
MVBench lab	63.7% ✎	▁··▃▄▁▃▄▄▄█▄ beats 3/26	medium · avg 67%	▇ 2.6%	blunt · at its level · small run

$ where to run it

Other ways to run it

Run it as published · needs ~8.29 GB, NVIDIA only · from github.com/vllm-project
vllm serve OpenGVLab/InternVL2-4B --trust-remote-code

Official weights and documentation: huggingface.co/OpenGVLab ↗

$ how the score was worked out · OpenGVLab InternVL2 model car… 53%, LMArena leaderboard dataset 22%, Microsoft Phi-3.5-vision-inst… 19% and 1 more

Each result implies a score on its own, after adjusting for how hard its test is and who ran it. The 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's added up by who ran the tests. How it works

source	results	share of the score	licence
OpenGVLab InternVL2 model cards (InternVL2-4B and InternVL2-26B), Evaluation tables · lab-reported	12	▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇ 53%	MIT
LMArena leaderboard dataset	1	▇▇▇▇▇▇▇ 22%	CC BY 4.0
Microsoft Phi-3.5-vision-instruct model card: Benchmarks table	5	▇▇▇▇▇▇ 19%	MIT
PhysBench: Benchmarking and Enhancing Vision-Language Models for Physical World Understanding (arXiv:2501.16411v2)	1	▇ 4%	CC-BY-4.0
prior · pull toward the middle		▇ 2%
weighted average		= 41.2

not counted (2)

OpenGVLab (Shanghai AI Laboratory) vendor-reported scores for its own InternVL models (InternVL2.5 report, arXiv 2412.05271v5)	CharXiv (reasoning questions) [correct]	24.5%	fewer than three models have this test yet, or this source ran it for this model alone
OpenGVLab (Shanghai AI Laboratory) vendor-reported scores for its own InternVL models (InternVL2.5 report, arXiv 2412.05271v5)	CharXiv descriptive questions [correct]	48.0%	fewer than three models have this test yet, or this source ran it for this model alone

$ improve this page

Update model information Add a place to run it Add a test result Propose a benchmark Correct a benchmark score Edit data on GitHub

Use a short GitHub form or edit the data through a pull request. Every number needs a source; here's which ones we can use.

data updated 2026-10-11 02:20 UTC · sources and licences · json