aya-vision-32b
by CohereLabs
33.1B params · image-text-to-text · 226 likes · 288 downloads
aya-vision-32b is a 33.1B parameter model. At Q4 quantization it requires 17GB of VRAM. It runs comfortably on GeForce RTX 4090 (39 tok/s), GeForce RTX 5090 (59 tok/s), A100 PCIe 80 GB (72 tok/s).
Inference providers
| Provider | $/1M in | $/1M out | Throughput |
|---|---|---|---|
| Cohere | 42 tok/s |
GPU compatibility
| GPU | VRAM | Q4 Decode | Verdict |
|---|---|---|---|
| GeForce RTX 4090 | 24GB | 39 tok/s | comfortable |
| GeForce RTX 5090 | 32GB | 59 tok/s | comfortable |
| M4 Max 128GB | 128GB | 14 tok/s | tight |
| M4 Pro 48GB | 48GB | 7 tok/s | tight |
| M4 Pro 24GB | 24GB | 7 tok/s | tight |
| A100 PCIe 80 GB | 80GB | 72 tok/s | comfortable |
| H100 SXM5 80 GB | 80GB | 146 tok/s | comfortable |
| GeForce RTX 3090 | 24GB | 35 tok/s | comfortable |
| Radeon RX 7900 XTX | 24GB | 28 tok/s | tight |
| M4 Max 64GB | 64GB | 14 tok/s | tight |