SpatialLM-Llama-1B
by manycore-research
1.6B params · text-generation · 993 likes · 253 downloads
SpatialLM-Llama-1B is a 1.6B parameter model. At Q4 quantization it requires 1GB of VRAM. It runs comfortably on GeForce RTX 4090 (821 tok/s), GeForce RTX 5090 (1231 tok/s), M4 Max 128GB (300 tok/s).
GPU compatibility
| GPU | VRAM | Q4 Decode | Verdict |
|---|---|---|---|
| GeForce RTX 4090 | 24GB | 821 tok/s | comfortable |
| GeForce RTX 5090 | 32GB | 1231 tok/s | comfortable |
| M4 Max 128GB | 128GB | 300 tok/s | comfortable |
| M4 Pro 48GB | 48GB | 150 tok/s | comfortable |
| M4 Pro 24GB | 24GB | 150 tok/s | comfortable |
| A100 PCIe 80 GB | 80GB | 1504 tok/s | comfortable |
| H100 SXM5 80 GB | 80GB | 3025 tok/s | comfortable |
| GeForce RTX 3090 | 24GB | 725 tok/s | comfortable |
| Radeon RX 7900 XTX | 24GB | 600 tok/s | comfortable |
| GeForce RTX 4080 | 16GB | 582 tok/s | comfortable |