作者: Jakub Rusinowski · 最后更新: 2026年7月21日
How much GPU VRAM you need to run Cosmos 3 Cosmos 3 Super by NVIDIA locally, a 64B-parameter model. Figures are quantized weights + KV cache + framework overhead from a memory-bandwidth roofline model.
Cosmos 3 Super needs about 39 GB VRAM at Q4_K_M.
| Quant | Bits/weight | Weights | Total VRAM |
|---|---|---|---|
| Q2_K | 2.63 | 21.0 GB | 21.8 GB |
| Q3_K_M | 3.41 | 27.3 GB | 28.1 GB |
| Q4_K_M | 4.83 | 38.6 GB | 39.4 GB |
| Q5_K_M | 5.67 | 45.4 GB | 46.2 GB |
| Q6_K | 6.56 | 52.5 GB | 53.3 GB |
| Q8_0 | 8.50 | 68.0 GB | 68.8 GB |
| F16 | 16.00 | 128.0 GB | 128.8 GB |
Switch quantization in the interactive calculator, or see the full Cosmos 3 model page.
Model creators: paste this into your Hugging Face model card README to link readers straight to this VRAM breakdown.
[](https://llmconfigurator.com/en/vram-calculator/cosmos-3-super)
Estimates only — actual VRAM varies with context length, batch size, runtime and KV-cache settings.