falcon-40b-instruct
by tiiuae
40.0B params · text-generation · 1.2k likes · 40.9k downloads
falcon-40b-instruct is a 40.0B parameter model. At Q4 quantization it requires 20GB of VRAM. It runs comfortably on GeForce RTX 4090 (32 tok/s), GeForce RTX 5090 (49 tok/s), A100 PCIe 80 GB (60 tok/s).
GPU compatibility
| GPU | VRAM | Q4 Decode | Verdict |
|---|---|---|---|
| GeForce RTX 4090 | 24GB | 32 tok/s | comfortable |
| GeForce RTX 5090 | 32GB | 49 tok/s | comfortable |
| M4 Max 128GB | 128GB | 12 tok/s | tight |
| M4 Pro 48GB | 48GB | 6 tok/s | tight |
| M4 Pro 24GB | 24GB | 6 tok/s | tight |
| A100 PCIe 80 GB | 80GB | 60 tok/s | comfortable |
| H100 SXM5 80 GB | 80GB | 120 tok/s | comfortable |
| GeForce RTX 3090 | 24GB | 29 tok/s | tight |
| Radeon RX 7900 XTX | 24GB | 24 tok/s | tight |
| M4 Max 64GB | 64GB | 12 tok/s | tight |