[whatmodel.sh](../../README.md) / Qwen2-VL 72B Instruct
# Qwen2-VL 72B Instruct
Alibaba · released 2024-09 · 73.4B params · tongyi-qianwen
**score 53.5** ±0.9 · #308 of 732 · from 29 benchmark results across 6 sources
Score within the [LMArena (text)](https://whatmodel.sh/?set=lmarena-text&score=set) comparison set only.
## What it's good at
● marks Qwen2-VL 72B Instruct among every model on that task or test, worst to best.
task vs the others rank counts tests
math ▁▁▃▅█●▄█▄▂▁▁ #244 of 431 16% 2
vision & documents ▁▁▂▂▁●▄▂▆▇█▄ #123 of 153 87% 18
math · #244 of 431 · 16% of its score
Beats [GPT-4o (2024-08-06)](gpt-4o-2024-08-06.md), [Command A (03-2025)](command-a.md), [GPT-4o (2024-05-13)](gpt-4o-2024-05-13.md) here, though they rank higher overall. Trails [Gemma 7B (base)](gemma-7b-base.md), [CodeGemma 7B (v1.0)](codegemma-7b.md), [Phi-2 (2.7B)](phi-2.md) here, though they rank lower overall.
test result vs the others in its run difficulty counts why
MathVista (testmini) (maker) 70.5% ▃··▃▅·▃●▅▅▅█ beats 5/16 medium · avg 69% 12% at its level · several runs
MATH-Vision (maker) 25.9% ▁▁·●··▄▁▃▇▇█ beats 2/24 hard · avg 25% 4.7% small run
vision & documents · #123 of 153 · 87% of its score
Beats [DeepSeek V4 Flash Vision (Exp)](deepseek-v4-flash-vision-exp.md), [GPT-4o mini (2024-07-18)](gpt-4o-mini-2024-07-18.md) here, though they rank higher overall. Trails [Qwen3.5 2B](qwen3-5-2b.md), [Qwen3-VL 4B Instruct](qwen3-vl-4b-instruct.md), [Gemma 4 E4B](gemma-4-e4b.md) here, though they rank lower overall.
test result vs the others in its run difficulty counts why
MMMU (val) (maker) 64.5% ▂···▆▂▂●▄▄▄█ beats 6/16 medium · avg 61% 16% at its level · several runs
MathVista (testmini) (maker) 70.5% ▃··▃▅·▃●▅▅▅█ beats 5/16 medium · avg 69% 12% at its level · several runs
MMStar (maker) 68.3% ▂··▄▂▂·●█▂▄▆ beats 6/16 medium · avg 65% 11% at its level · several runs
MMMU Pro (maker) 46.2% ▃·█·▅●▃▃▅▅▃█ beats 6/16 medium · avg 46% 11% at its level · several runs
Arena vision (style controlled) 1084.6 ▁▁▂▂▂●▃▄▅▅█▇ beats 27/152 avg 1061 6.1% blunt
MATH-Vision (maker) 25.9% ▁▁·●··▄▁▃▇▇█ beats 2/24 hard · avg 25% 4.7% small run
CapArena-Auto detailed image capt… -9.0000 ▃▃▃·▃●·▅▃▃▃▃ beats 5/12 avg -4.498 3.9% typical
DocVQA (maker) 96.5% ▃··▅▃▅▃█▃█●▅ beats 14/17 very easy · avg 94% 3.7% blunt · maxed out · several runs
HallusionBench (maker) 58.1% ▃·▃·▅▃▅●▃▅█▅ beats 8/16 medium · avg 54% 3.6% blunt · at its level
ChartQA (maker) 88.3% ▂··▂·▂▃·▅●▆█ beats 8/17 easy · avg 85% 3.0% blunt · several runs
MM-MT-Bench (Mistral, GPT-4o judg… 6.5900 MM-MT-Bench score (1-10) ▅··▃·█·▃●··▅ beats 7/9 avg 5.605 MM-MT-Bench s… 2.8% sharp · noisy
RealWorldQA (maker) 77.8% ▅▃·▃▅▅·█·▃●█ beats 13/16 easy · avg 70% 2.6% blunt · at its level
InfographicVQA (maker) 84.5% ▃·▃▅█▃▃▃●▃▃▃ beats 10/13 easy · avg 80% 2.5% small run
VQAv2 (val) 81.6% ▃···█▃··▅▃▃● beats 9/9 easy · avg 73% 1.5% blunt
OCRBench (maker) 87.7% ▅█·▃▃█●▅▅▃·▃ beats 10/16 easy · avg 84% 0.9% blunt
Ai2 Molmo human preference Elo (S… 1037.0 ▄▄·▄·▄▄●·▄·█ beats 5/8 avg 1032 0.9% blunt
CC-OCR (maker) 68.7% ▁·●··▂▂▃▃▂▆█ beats 2/26 easy · avg 72% 0.7% blunt · at its level · small run
TextVQA (maker) 85.5% █······▄·▄▄● beats 5/5 easy · avg 80% 0.5% blunt
other tests (3)
test result vs the others in its run difficulty counts why
Video-MME (without subtitles) (ma… 71.2% ▃··█··▅●▃▃▅▃ beats 6/13 medium · avg 66% 6.6% blunt · at its level · several runs
Video-MME (with subtitles) (maker) 77.8% ▁▁▁▁▁·▁●▁▁▁█ beats 6/16 medium · avg 73% 4.2% at its level · small run · several runs
MVBench (maker) 73.6% ▂▂▂·▂█··▆●·▂ beats 12/13 medium · avg 67% 2.0% blunt · at its level
source results share of the score
Qwen (submitted / maker-reported) 15 69% ▇▇▇▇▇▇▇▇
Qwen2.5-VL-32B-Instruct model card (sub… 6 11% ▇
Pixtral 12B 5 8% ▇
LMArena leaderboard dataset 1 6% ▇
CapArena: Benchmarking and Analyzing De… 1 4% ▇
Ai2 Molmo-72B-0924 model card: Human Pr… 1 1% ▇
prior: pull toward the middle 0% ▇
weighted average = 53.5
Licences: Qwen2.5-VL-32B-Instruct model card ([Apache-2.0](https://www.apache.org/licenses/LICENSE-2.0)) · Pixtral 12B ([CC BY 4.0](http://creativecommons.org/licenses/by/4.0/)) · 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/)) · Ai2 Molmo-72B-0924 model card: Human Preference Elo Rating column ([Apache-2.0](https://huggingface.co/allenai/Molmo-72B-0924/blob/ab586dd9640de059566975b8e5bcd0c6fdc73956/README.md))
## How to run it
### llama.cpp (recommended)
1. Install llama.cpp. Homebrew and winget installs name the commands llama-cli and llama-server instead. [Other ways to install](https://github.com/ggml-org/llama.cpp/blob/master/docs/install.md)
```sh
# macOS / Linux
curl -LsSf https://llama.app/install.sh | sh
# Windows
irm https://llama.app/install.ps1 | iex
```
2. Run Qwen2-VL 72B Instruct (Q4_K_M, needs ~48.82 GB, from [huggingface.co/bartowski](https://huggingface.co/bartowski/Qwen2-VL-72B-Instruct-GGUF/tree/2b2dc80ac1a8727d8d90f86308d849318c8ef311))
```sh
llama cli -hf bartowski/Qwen2-VL-72B-Instruct-GGUF:Q4_K_M
```
3. Or start a local server and open the address it prints (http://localhost:9931)
```sh
llama serve -hf bartowski/Qwen2-VL-72B-Instruct-GGUF:Q4_K_M
```
Other versions: Q2_K (~31.21 GB) `llama cli -hf bartowski/Qwen2-VL-72B-Instruct-GGUF:Q2_K`
### LM Studio
1. Install LM Studio. Open the app once before using the lms command. [Download it (macOS / Windows / Linux)](https://lmstudio.ai/download)
2. Run Qwen2-VL 72B Instruct (Q4_K_M, needs ~48.82 GB, from [huggingface.co/bartowski](https://huggingface.co/bartowski/Qwen2-VL-72B-Instruct-GGUF/tree/2b2dc80ac1a8727d8d90f86308d849318c8ef311))
```sh
lms get https://huggingface.co/bartowski/Qwen2-VL-72B-Instruct-GGUF@Q4_K_M -y
```
3. Then run lms ls to see its name, and lms chat followed by that name
Other versions: Q2_K (~31.21 GB) `lms get https://huggingface.co/bartowski/Qwen2-VL-72B-Instruct-GGUF@Q2_K -y` · 4bit (~41.3 GB, Mac only) `lms get https://huggingface.co/mlx-community/Qwen2-VL-72B-Instruct-4bit -y` · 8bit (~78 GB, Mac only) `lms get https://huggingface.co/mlx-community/Qwen2-VL-72B-Instruct-8bit -y`
### Ollama
1. Install Ollama. [Download it (macOS / Windows)](https://ollama.com/download)
```sh
# Linux
curl -fsSL https://ollama.com/install.sh | sh
```
2. Run Qwen2-VL 72B Instruct (Q2_K, needs ~31.21 GB, from [huggingface.co/bartowski](https://huggingface.co/bartowski/Qwen2-VL-72B-Instruct-GGUF/tree/2b2dc80ac1a8727d8d90f86308d849318c8ef311))
```sh
ollama run hf.co/bartowski/Qwen2-VL-72B-Instruct-GGUF:Q2_K
```
### MLX (Mac)
1. Install MLX (Mac). Apple silicon Macs only, text only.
```sh
# with uv
uv tool install mlx-lm
# or
pip install mlx-lm
```
2. Run Qwen2-VL 72B Instruct (4bit, needs ~41.3 GB, Mac only, from [huggingface.co/mlx-community](https://huggingface.co/mlx-community/Qwen2-VL-72B-Instruct-4bit/tree/7c750d8a653fc3ea1ab19b70564b21b25e7c78d8))
```sh
mlx_lm.chat --model mlx-community/Qwen2-VL-72B-Instruct-4bit
```
3. Or start an OpenAI-compatible server at http://localhost:8080/v1
```sh
mlx_lm.server --model mlx-community/Qwen2-VL-72B-Instruct-4bit
```
Other versions: 8bit (~78 GB, Mac only) `mlx_lm.chat --model mlx-community/Qwen2-VL-72B-Instruct-8bit`
Model repository and documentation: [huggingface.co/Qwen](https://huggingface.co/Qwen/Qwen2-VL-72B-Instruct)
[interactive version](https://whatmodel.sh/model/qwen2-vl-72b) · [JSON](https://whatmodel.sh/api?format=json&all=1)