[whatmodel.sh](../../README.md) / Aya Vision 32B # Aya Vision 32B Cohere · released 2025-03 · 33.1B params · cc-by-nc-4.0 **score 43.5** ±2.9 · #509 of 732 · from 3 benchmark results across 3 sources Score within the [LMArena (text)](https://whatmodel.sh/?set=lmarena-text&score=set) comparison set only. ## What it's good at ● marks Aya Vision 32B among every model on that task or test, worst to best. 
task                vs the others         rank  counts  tests
vision & documents  ▁▁●▂▁▃▄▂▆▇█▄   #141 of 153     96%      3
vision & documents · #141 of 153 · 96% of its score Beats [Gemma 3 4B](gemma-3-4b.md), [LLaVA-OneVision Qwen2 72B (ov-sft)](llava-onevision-qwen2-72b-ov.md), [CogVLM2 Llama3 Chat 19B](cogvlm2-llama3-chat-19b.md) here, though they rank higher overall. Trails [Qwen3.5 0.8B](qwen3-5-0-8b.md) here, though they rank lower overall. 
test                                result  vs the others in its run   difficulty           counts  why
Arena vision (style controlled)     1000.4  ▁▁▂●▂▂▃▄▅▅█▇ beats 11/152  avg 1061                58%  blunt
Kaleidoscope (in-language multimo…   39.3%  █·●···█·█·██ beats 1/5     medium · avg 46%        30%  at its level
VMMU (Vietnamese multimodal under…   22.5%  ●······▃·▅·█ beats 1/7     very hard · avg 36%    7.9%  too hard for it · noisy
## 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
LMArena leaderboard dataset                     1   58% ▇▇▇▇▇▇▇
Kaleidoscope: In-language Exams for Mas…        1   30% ▇▇▇▇
VMMU: A Vietnamese Multitask Multimodal…        1    8% ▇
prior: pull toward the middle                        4% ▇
weighted average                                   = 43.5
Licences: LMArena leaderboard dataset ([CC BY 4.0](https://creativecommons.org/licenses/by/4.0/)) · Kaleidoscope: In-language Exams for Massively Multilingual Vision Evaluation ([CC-BY-4.0](https://creativecommons.org/licenses/by/4.0/)) · VMMU: A Vietnamese Multitask Multimodal Understanding and Reasoning Benchmark ([CC-BY-4.0](https://creativecommons.org/licenses/by/4.0/)) ## How to run it ### Through an API api (Cohere) · price unknown · [docs.cohere.com](https://docs.cohere.com/docs/models) [context source](https://huggingface.co/CohereLabs/aya-vision-32b/blob/0554d66834922fc0f2e5f47a12f78464f4a98533/README.md): Context length stated on the official model card. ### Other ways to run it 1. Run it as published (fp16, needs ~66.28 GB, from [huggingface.co/CohereLabs](https://huggingface.co/CohereLabs/aya-vision-32b)) ```sh pip install 'git+https://github.com/huggingface/transformers.git@v4.49.0-AyaVision' torch accelerate && python -c "import torch; from transformers import AutoProcessor, AutoModelForImageTextToText; r='CohereLabs/aya-vision-32b'; p=AutoProcessor.from_pretrained(r); m=AutoModelForImageTextToText.from_pretrained(r, device_map='auto', torch_dtype=torch.float16); i=p.apply_chat_template([{'role': 'user', 'content': [{'type': 'image', 'url': 'https://pbs.twimg.com/media/Fx7YvfQWYAIp6rZ?format=jpg&name=medium'}, {'type': 'text', 'text': 'What does the text in the image say?'}]}], padding=True, add_generation_prompt=True, tokenize=True, return_dict=True, return_tensors='pt').to(m.device); g=m.generate(**i, max_new_tokens=300, do_sample=True, temperature=0.3); print(p.tokenizer.decode(g[0][i['input_ids'].shape[-1]:], skip_special_tokens=True))" ``` Model repository and documentation: [huggingface.co/CohereLabs](https://huggingface.co/CohereLabs/aya-vision-32b) [interactive version](https://whatmodel.sh/model/aya-vision-32b) · [JSON](https://whatmodel.sh/api?format=json&all=1)