Model
Model explorer

Aya Vision 32B

OPEN
Cohere Labs · Aya Vision family · released Mar 4, 2025

Cohere's first vision-language model, pairing an Aya Expanse 32B backbone with a SigLIP2 encoder for multilingual multimodal tasks.

ReasoningCodingVisionFunction callingTool useAgentic
713.9
Elo · rank #262
Parameters
32B
Active params
32B (dense)
Context
16K tokens
Architecture
Aya Expanse 32B decoder + SigLIP2-patch14-384 vision encoder
License
CC-BY-NC
Languages
23+
API price (in/out)
No hosted API
Modalities
text · vision
Benchmark results
Bar shows position within the tracked field; marker = field best
GPQA DiamondReasoning34.4%#242
best: GPT-5.6 · 94.6%
HumanEvalCoding62.2%#97
best: Claude Opus 4.5 · 99.4%
IFEvalReasoning78.5%#94
best: Gemma 4 26B A4B · 98.5%
MATH-500Math69.3%#95
best: GPT-5 · 99.4%
MBPPCoding70.4%#36
best: Llama-3.3-Nemotron-Super-49B v1 (Reasoning On) · 91.3%
MGSMMath79.5%#28
best: OpenAI o4-mini · 93.7%
MMLU-ProKnowledge47.2%#138
best: Claude Fable 5 · 91.5%
MMLUKnowledge72.1%#136
best: OpenAI o3 · 92.9%
SimpleQAKnowledge7.7%#42
best: GPT-4.5 · 62.5%
Run it locally
VRAM @ Q4
—
VRAM @ FP16
66 GB
Fits on (Q4)
Multi-node cluster required
Throughput data unavailable.
Quantizations
—
Aya Vision family
Elo progression across releases
Aya Vision 32BMar 2025713.9
Resources
HFCohereLabs/aya-vision-32bDOCSdocs.cohere.com/docs/aya-visionAPIAvailable via Cohere Labs API
API price weights · each benchmark row carries its own source badge (see methodology)