whatmodel.sh / Aya Vision 32B

Aya Vision 32B

+ contribute
Update model informationAdd a place to run itAdd a test resultPropose a benchmarkCorrect a benchmark scoreEdit data on GitHub

Aya Vision 32B is an open-weights language model from Cohere, released March 2025. Its published size is 33.1B parameters and a 16k-token context window. Licence: cc-by-nc-4.0.

Catalogue identity notes

Arena entry 'c4ai-aya-vision-32b' (vision). LMArena's leaderboard table (arena-leaderboard space @6f2eeeb) labels 'c4ai-aya-vision-32b' as 'Aya-Vision-32B' (Cohere, CC-BY-NC-4.0), linking https://huggingface.co/CohereForAI/aya-vision-32b. Maker repo CohereLabs/aya-vision-32b @0554d66: licence tag cc-by-nc-4.0. 33,137,896,688 parameters from the Hugging Face safetensors metadata (dense, so active = total). Released 2025-03: month the maker's Hugging Face repo was created (createdAt 2025-03-02). Cohere models page lists c4ai-aya-vision-32b as Live.

42.4±3.0#495 of 719 LLMsfrom 3 benchmark results across 3 sources

$ how it compares

The LLMs ranked either side of Aya Vision 32B, out of 828 with a score, then more from Cohere.

#modelscorewhereRAMreleased
583Phi-3-small-8k-instruct Microsoft43±2local~17GB2024-05
584CodeGemma 7B IT (v1.0) Google43±1local~7.1GB2024-03
585Aya Vision 32B Cohere42±3api price unknown +local~71GB2025-03
586Tulu 2 DPO 70B Ai242±1local~32GB2023-11
587Mathstral 7B v0.1 Mistral AI42±1local~5.8GB2024-07
more from Cohere
76North Small Translate 1.0 Cohere71±6api price unknown2026-08
175Command A+ (05-2026) Cohere64±1api price unknown +local~71GB2026-05
277Command A (03-2025) Cohere57±1api $4.38/M +local~37GB2025-03

see all LLMs · find the ones that fit your machine

$ what it's good at

vision & documents▁▁▃▂▁▃▄▂▆▇█▄ #135 of 147 · 96% of its score · 3 tests

Beats LLaVA-OneVision Qwen2 72B (ov-sft), Gemma 3 4B, CogVLM2 Llama3 Chat 19B here, though they rank higher overall. Trails Qwen3.5 0.8B here, though they rank lower overall.

testresultvs the others in its rundifficultycountswhy
Arena vision (style controlled)1000.2 ✎▁▁▂▃▂▂▄▅▅▆██ beats 11/145avg 1068▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇ 57%blunt
Kaleidoscope (in-language multimodal exams, 18 languages, overall accuracy)39.3% ✎█·█···█·█·██ beats 1/5medium · avg 47%▇▇▇▇▇▇▇▇▇▇▇▇ 31%at its level
VMMU (Vietnamese multimodal understanding and reasoning, mean accuracy over 7 tasks)22.5% ✎▅······▃·▅·█ beats 1/7very hard · avg 36%▇▇▇ 8.7%too hard for it · noisy

$ where to run it

Through an API

api · price unknown · 16k context · docs.cohere.com ↗

Other ways to run it

  1. Run it as published · needs ~66.28 GB · from huggingface.co/CohereLabs
    pip install 'git+https://github.com/huggingface/[email protected]' 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))"

Weights and documentation: huggingface.co/CohereLabs ↗

$ how the score was worked out · LMArena leaderboard dataset 57%, Kaleidoscope: In-language Exa… 31%, VMMU: A Vietnamese Multitask… 9%

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

sourceresultsshare of the scorelicence
LMArena leaderboard dataset1▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇ 57%CC BY 4.0
Kaleidoscope: In-language Exams for Massively Multilingual Vision Evaluation (arXiv:2504.07072v2)1▇▇▇▇▇▇▇▇▇ 31%CC-BY-4.0
VMMU: A Vietnamese Multitask Multimodal Understanding and Reasoning Benchmark (arXiv:2508.13680v5)1▇▇▇ 9%CC-BY-4.0
prior · pull toward the middle▇ 4%
weighted average= 42.4

$ improve this page

Update model informationAdd a place to run itAdd a test resultPropose a benchmarkCorrect a benchmark scoreEdit 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 · sources and licences · json