CohereLanguage model
Aya Vision 32B
Aya Vision 32B has no overall index value yet because there are too few independent benchmark results.
20out of 100same as last month
based on search, usage and attention
Summary
Aya Vision 32B extends Cohere's multilingual Aya approach with visual inputs. The open vision-language model accepts text and images for OCR, image captioning, visual reasoning, summarization, visual question answering, and code from images. Its vision encoder is based on SigLIP2 and connects to the Aya Expanse language model through a multimodal adapter.
The model card names 23 languages, 32 billion parameters, and up to twelve image tiles plus a thumbnail per request. Output remains text-based. Aya Vision 8B is the smaller multimodal family variant.
All benchmark results
Every published result with its source, test setup and comparison models.
Show all 84 published benchmark results
Page 1 of 7
Measurements without matching peer values
These measurements have no matching peer values under the same test conditions. Their original values and sources remain available here.
AyaVisionBench win rate vs Llama 3.2 90B Vision (Show measurement, test conditions, and source)
- Source value
- 63.9
- Score
- 63.9
- Metric
- pairwise win rate
- Unit
- %
- Benchmark version
- AyaVisionBench; 23-language average
- Category
- multimodal
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Cohere for AI Aya Vision win-rate figure
- Status
- active
- Retrieved at
- 2026-10-15
- Methodology
- Vendor-reported result. Exact Aya Vision 32B release model. The figure compares pairwise win, loss, and tie rates on AyaVisionBench; the article says results are averaged across 23 languages. The card identifies Claude 3.7 Sonnet as the judge. Prompt details and the number of judgments are not stated.
AyaVisionBench win rate vs Molmo 72B (Show measurement, test conditions, and source)
- Source value
- 58.9
- Score
- 58.9
- Metric
- pairwise win rate
- Unit
- %
- Benchmark version
- AyaVisionBench; 23-language average
- Category
- multimodal
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Cohere for AI Aya Vision win-rate figure
- Status
- active
- Retrieved at
- 2026-10-15
- Methodology
- Vendor-reported result. Exact Aya Vision 32B release model. The figure compares pairwise win, loss, and tie rates on AyaVisionBench; the article says results are averaged across 23 languages. The card identifies Claude 3.7 Sonnet as the judge. Prompt details and the number of judgments are not stated.
AyaVisionBench win rate vs Qwen2.5-VL 72B (Show measurement, test conditions, and source)
- Source value
- 50.2
- Score
- 50.2
- Metric
- pairwise win rate
- Unit
- %
- Benchmark version
- AyaVisionBench; 23-language average
- Category
- multimodal
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Cohere for AI Aya Vision win-rate figure
- Status
- active
- Retrieved at
- 2026-10-15
- Methodology
- Vendor-reported result. Exact Aya Vision 32B release model. The figure compares pairwise win, loss, and tie rates on AyaVisionBench; the article says results are averaged across 23 languages. The card identifies Claude 3.7 Sonnet as the judge. Prompt details and the number of judgments are not stated.
AyaVisionBench win rate vs Pangea 7B (Show measurement, test conditions, and source)
- Source value
- 79.1
- Score
- 79.1
- Metric
- pairwise win rate after model-scale stage
- Unit
- %
- Benchmark version
- AyaVisionBench; 23-language average
- Category
- multimodal
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Cohere for AI Aya Vision model-scale win-rate figure
- Status
- active
- Retrieved at
- 2026-10-15
- Methodology
- Vendor-reported result. The figure's final '+Model Scale' stage is the publisher's 8B-to-32B scaling step. Cohere describes this as AyaVisionBench win rate against Pangea 7B, averaged across 23 languages. The figure explicitly prints 79.1%; no bar estimate is used.
G-MMLU Lite (Show measurement, test conditions, and source)
- Source value
- 63.58
- Score
- 63.58
- Metric
- accuracy
- Unit
- %
- Category
- language-understanding
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Aya Vision technical report, arXiv:2505.08751v1
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Vendor-reported result. Aya Vision 32B, paper text-only benchmark table. The table does not state a prompt-shot count.
MGSM (Show measurement, test conditions, and source)
- Source value
- 79.46
- Score
- 79.46
- Metric
- accuracy
- Unit
- %
- Category
- mathematical-reasoning
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Aya Vision technical report, arXiv:2505.08751v1
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Vendor-reported result. Aya Vision 32B, paper text-only benchmark table. The table does not state a prompt-shot count.
IFEval (Show measurement, test conditions, and source)
- Source value
- 78.5
- Score
- 78.5
- Metric
- accuracy
- Unit
- %
- Category
- language-understanding
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Aya Vision technical report, arXiv:2505.08751v1
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Vendor-reported result. Aya Vision 32B, paper text-only benchmark table. The table does not state a prompt-shot count.
m-ArenaHard win rate vs Llama 3.2 90B Vision (Show measurement, test conditions, and source)
- Source value
- 43.2
- Score
- 43.2
- Metric
- pairwise win rate
- Unit
- %
- Category
- open-ended-generation
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Aya Vision technical report, arXiv:2505.08751v1
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Vendor-reported result. m-ArenaHard, pairwise text-only evaluation, average across 23 languages; GPT-4o-2024-11-20 judge.
m-ArenaHard win rate vs Molmo 72B (Show measurement, test conditions, and source)
- Source value
- 77.3
- Score
- 77.3
- Metric
- pairwise win rate
- Unit
- %
- Category
- open-ended-generation
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Aya Vision technical report, arXiv:2505.08751v1
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Vendor-reported result. m-ArenaHard, pairwise text-only evaluation, average across 23 languages; GPT-4o-2024-11-20 judge.
m-ArenaHard win rate vs Qwen 2.5 VL 72B (Show measurement, test conditions, and source)
- Source value
- 50.9
- Score
- 50.9
- Metric
- pairwise win rate
- Unit
- %
- Category
- open-ended-generation
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Aya Vision technical report, arXiv:2505.08751v1
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Vendor-reported result. m-ArenaHard, pairwise text-only evaluation, average across 23 languages; GPT-4o-2024-11-20 judge.
Cedar accuracy, multimodal (reported Avg.) (Show measurement, test conditions, and source)
- Source value
- 39.4
- Score
- 39.4
- Metric
- standard accuracy
- Unit
- %
- Category
- multilingual-cultural-reasoning
- Direction
- Higher is better
- Source type
- independent-evaluator
- Evaluator
- Cedar independent evaluation, arXiv:2601.13024v1
- Status
- disputed
- Retrieved at
- 2026-10-17
- Methodology
- Independent evaluation. Aya Vision 32B exact CohereLabs checkpoint, Cedar Table 1 Avg. column, multimodal subset; paper defines Avg. as macro-average across languages.
- Context
- The paper calls the aggregate a macro-average across languages, but the published aggregate differs from the mean of the displayed non-missing language cells. Kept as disputed, not a reconciled score.
Cedar accuracy, text-only (reported Avg.) (Show measurement, test conditions, and source)
- Source value
- 43.88
- Score
- 43.88
- Metric
- standard accuracy
- Unit
- %
- Category
- multilingual-cultural-reasoning
- Direction
- Higher is better
- Source type
- independent-evaluator
- Evaluator
- Cedar independent evaluation, arXiv:2601.13024v1
- Status
- disputed
- Retrieved at
- 2026-10-17
- Methodology
- Independent evaluation. Aya Vision 32B exact CohereLabs checkpoint, Cedar Table 1 Avg. column, text-only subset; paper defines Avg. as macro-average across languages.
- Context
- The paper calls the aggregate a macro-average across languages, but the published aggregate differs from the mean of the displayed non-missing language cells. Kept as disputed, not a reconciled score.
Page 1 of 7
Popularity compared
How often the model is searched for, used and mentioned, compared with every model of its category.
Popularity index
The 100 most popular models of this category, 0 to 100, higher is more popular
20 of 101 models
As of November 2026 · Source: gradually.ai popularity index
Over time
Weekly values from October 18, 2026 to November 7, 2026
Aya Vision 32BCommand A Translate 20Codestral-22B-v0.1 19DBRX 19Llama 3.1 70B 19
Method changes in the timeline, a jump at these points is no market event:
- 1November 1, 2026: Value only with search or usage data, 30-day attention
What is gaining popularity right now, across models, vendors, harnesses and AI subscriptions, is collected in All AI trends.
Similar models
More models of the same kind as Aya Vision 32B, chosen by model line, developer, size, price and release date.
- Aya Expanse 32BCohere · same developer
- North Mini CodeCohere · same developer
- Command ACohere · same developer
- Aya Vision 8BCohere · same developer
- Command R7B ArabicCohere · same developer
- Command A VisionCohere · same developer
- Aya Expanse 8BCohere · same developer
- Command A ReasoningCohere · same developer
All specifications
Architecture
Input and output
Watermarking and provenance
AI watermarking refers to invisible signals embedded in the output. Visible marks and provenance metadata are listed separately. Unknown means there is insufficient evidence for that particular field.
- AI watermarking
- Unknown
- no source
- Watermarking technique
- Not known
- no source
Licence and access
Frequently asked questions about Aya Vision 32B
How large is the context window of Aya Vision 32B?
16,000 tokens. That is how much text Aya Vision 32B reads at once, about 20 A4 pages.
Page count as a rule of thumb from Artificial Analysis
Can I run Aya Vision 32B locally?
Yes, the weights are open.
Stated by Cohere
- Open weights
- yes
- License
- CC BY-NC-4.0
- Parameters
- 32B
- active per token
- n/a
Sources and data date
Show all 31 sources
Benchmark sources at a glance
Which independent evaluators tested this model. This is coverage, not a quality ranking.
No active result from the reviewed external evaluators can currently be attributed to this model. Other source evidence remains available below.
Without vendor claims and historical or disputed results; those stay labelled in the single values.