Phi-3.5-MoE-instruct
by microsoft
41.9B params · text-generation · 571 likes · 89.0k downloads
Phi-3.5-MoE-instruct is a 41.9B parameter model. At Q4 quantization it requires 21GB of VRAM. It runs comfortably on GeForce RTX 5090 (47 tok/s), A100 PCIe 80 GB (57 tok/s), H100 SXM5 80 GB (115 tok/s).
GPU compatibility
| GPU | VRAM | Q4 Decode | Verdict |
|---|---|---|---|
| GeForce RTX 5090 | 32GB | 47 tok/s | comfortable |
| M4 Max 128GB | 128GB | 11 tok/s | tight |
| M4 Pro 48GB | 48GB | 5 tok/s | tight |
| A100 PCIe 80 GB | 80GB | 57 tok/s | comfortable |
| H100 SXM5 80 GB | 80GB | 115 tok/s | comfortable |
| M4 Max 64GB | 64GB | 11 tok/s | tight |
| H200 SXM 141 GB | 141GB | 168 tok/s | comfortable |
| M2 Ultra 192GB | 192GB | 16 tok/s | tight |