Llama 3.1 8B: how much VRAM?
In Q4_K_M with 8k tokens of context, Llama 3.1 8B needs about 6 GB of memory. The cheapest option: RTX 3060 12 GB (€1,180 for the complete machine).
Model sheet
ollama run llama3.1:8bWeights on Hugging Face: meta-llama/Llama-3.1-8B-Instruct
VRAM by precision and context
| Precision | 4k | 8k | 32k | 128k |
|---|---|---|---|---|
| FP16 | 18 GB | 18 GB | 21 GB | 33 GB |
| FP8 | 9 GB | 10 GB | 13 GB | 25 GB |
| Q8_0 | 10 GB | 10 GB | 13 GB | 25 GB |
| Q6_K | 8 GB | 8 GB | 11 GB | 23 GB |
| Q4_K_M | 6 GB | 6 GB | 9 GB | 21 GB |
| Q3_K_M | 5 GB | 6 GB | 9 GB | 21 GB |
One user. Each simultaneous user adds their own context cache.
Frequently asked questions
How much VRAM does Llama 3.1 8B need?
With an 8k-token context and one user, Llama 3.1 8B needs about 6 GB in Q4, 10 GB in Q8 and 18 GB in FP16.
What is the cheapest hardware to run Llama 3.1 8B?
In Q4, the cheapest option in our catalogue to run Llama 3.1 8B is RTX 3060 12 GB, about €1,180 for the complete machine.
How fast does Llama 3.1 8B run?
Expect around 40 tokens per second for one user on RTX 3060 12 GB in Q4 (estimate based on memory bandwidth).
On which hardware?
| Hardware | Cards | Speed | Machine |
|---|---|---|---|
| RTX 3060 12 GB | 1 is enough | ~40 tok/s | €1,180 |
| RTX 4060 Ti 16 GB | 1 is enough | ~32 tok/s | €1,350 |
| Radeon RX 7800 XT 16 GB | 1 is enough | ~69 tok/s | €1,400 |
| RTX 4070 12 GB | 1 is enough | ~56 tok/s | €1,450 |
| RTX 3090 24 GB (used) | 1 is enough | ~104 tok/s | €1,650 |
| RTX 4070 Ti SUPER 16 GB | 1 is enough | ~75 tok/s | €1,750 |
| Radeon RX 7900 XTX 24 GB | 1 is enough | ~107 tok/s | €1,850 |
| RTX 4080 SUPER 16 GB | 1 is enough | ~82 tok/s | €1,950 |
| RTX 5080 16 GB | 1 is enough | ~107 tok/s | €2,100 |
| Mini-PC Ryzen AI Max+ 395 128 GB | 1 is enough | ~28 tok/s | €2,300 |
| RTX 4090 24 GB | 1 is enough | ~112 tok/s | €2,900 |
| RTX 5090 32 GB | 1 is enough | ~199 tok/s | €3,500 |
| NVIDIA DGX Spark 128 GB | 1 is enough | ~30 tok/s | €4,000 |
| Mac Studio M4 Max 128 GB | 1 is enough | ~61 tok/s | €4,400 |
| RTX 6000 Ada 48 GB | 1 is enough | ~107 tok/s | €7,900 |
| RTX PRO 6000 Blackwell 96 GB | 1 is enough | ~199 tok/s | €9,900 |
| Mac Studio M3 Ultra 512 GB | 1 is enough | ~91 tok/s | €11,500 |
| L40S 48 GB | 1 is enough | ~96 tok/s | €13,500 |
| A100 80 GB | 1 is enough | ~227 tok/s | €17,000 |
| H100 80 GB | 1 is enough | ~372 tok/s | €34,000 |
| H200 141 GB | 1 is enough | ~533 tok/s | €39,000 |
| B200 192 GB | 1 is enough | ~889 tok/s | €48,000 |
Estimated complete machine: cards + host PC (€900 for 1–2 consumer cards, €2,800 beyond, a €6,000 server for datacenter cards). Macs and mini-PCs are sold complete.
Similar models
From: RTX 3060 12 GB · €1,180
From: RTX 3060 12 GB · €1,180
From: RTX 3060 12 GB · €1,180
From: Mini-PC Ryzen AI Max+ 395 128 GB · €2,300