tiny-aya-water
by CohereLabs
3.3B params · text-generation · 26 likes · 952 downloads
tiny-aya-water is a 3.3B parameter model. At Q4 quantization it requires 2GB of VRAM. It runs comfortably on GeForce RTX 4090 (392 tok/s), GeForce RTX 5090 (587 tok/s), M4 Max 128GB (143 tok/s).
Inference providers
| Provider | $/1M in | $/1M out | Throughput |
|---|---|---|---|
| Cohere | 100 tok/s |
GPU compatibility
| GPU | VRAM | Q4 Decode | Verdict |
|---|---|---|---|
| GeForce RTX 4090 | 24GB | 392 tok/s | comfortable |
| GeForce RTX 5090 | 32GB | 587 tok/s | comfortable |
| M4 Max 128GB | 128GB | 143 tok/s | comfortable |
| M4 Pro 48GB | 48GB | 71 tok/s | comfortable |
| M4 Pro 24GB | 24GB | 71 tok/s | comfortable |
| A100 PCIe 80 GB | 80GB | 718 tok/s | comfortable |
| H100 SXM5 80 GB | 80GB | 1444 tok/s | comfortable |
| GeForce RTX 3090 | 24GB | 346 tok/s | comfortable |
| Radeon RX 7900 XTX | 24GB | 286 tok/s | comfortable |
| GeForce RTX 4080 | 16GB | 278 tok/s | comfortable |