GPT-NeoXT-Chat-Base-20B
by togethercomputer
27.2B params · text-generation · 694 likes · 888 downloads
GPT-NeoXT-Chat-Base-20B is a 27.2B parameter model. At Q4 quantization it requires 14GB of VRAM. It runs comfortably on GeForce RTX 4090 (48 tok/s), GeForce RTX 5090 (72 tok/s), A100 PCIe 80 GB (88 tok/s).
GPU compatibility
| GPU | VRAM | Q4 Decode | Verdict |
|---|---|---|---|
| GeForce RTX 4090 | 24GB | 48 tok/s | comfortable |
| GeForce RTX 5090 | 32GB | 72 tok/s | comfortable |
| M4 Max 128GB | 128GB | 17 tok/s | tight |
| M4 Pro 48GB | 48GB | 8 tok/s | tight |
| M4 Pro 24GB | 24GB | 8 tok/s | tight |
| A100 PCIe 80 GB | 80GB | 88 tok/s | comfortable |
| H100 SXM5 80 GB | 80GB | 177 tok/s | comfortable |
| GeForce RTX 3090 | 24GB | 42 tok/s | comfortable |
| Radeon RX 7900 XTX | 24GB | 35 tok/s | comfortable |
| GeForce RTX 4080 | 16GB | 34 tok/s | comfortable |