What Do You Need to Run Llama Locally? VRAM by Version
Llama currently has 4 versions you can deploy locally. The smallest, Llama 3.2 3B, needs 4GB of VRAM at Q4; the largest, Llama 4 Scout 17B 16E, needs 10GB. The cheapest device that gets one running is the Tesla V100 16GB, at about $305.
VRAM needed for each Llama version
eBay
| Version | Released | Context | |||
|---|---|---|---|---|---|
| Llama 3.2 3BDense | 3.2B | Sep 18, 2024 | 128K | 4 | Tesla V100 16GB$305 · 16 GB |
| Llama 3.1 8BDense | 8B | Jul 18, 2024 | 128K | 7 | Tesla V100 16GB$305 · 16 GB |
| Llama 3.3 70BDense | 70.6B | Nov 26, 2024 | 128K | 43 | Radeon PRO W7900$3,495 · 48 GB |
| Llama 4 Scout 17B 16EMoE | 108.6B (17B active) | Apr 2, 2025 | 10M | 10plus 55 GB of RAM | Tesla V100 16GB$305 · 16 GB · Experts offloaded |
Minimum VRAM = Q4_K_M weights + KV cache for an 8K context + 1GB runtime overhead, rounded up. The cheapest device is the first one that fits when sorting by median price on the current price basis.
About Llama
Llama is published by Meta AI under Llama 3.2 Community License / Llama 3.1 Community License / Llama 3.3 Community License / Llama 4 Community License. This page covers 4 versions ranging from 3.2B to 108.6B parameters. New official releases are scanned weekly; architecture parameters and sourced or estimated weight sizes are reviewed before publication.