Llama 3.3 70B: how much VRAM?
In Q4_K_M with 8k tokens of context, Llama 3.3 70B needs about 47 GB of memory. The cheapest option: Mini-PC Ryzen AI Max+ 395 128 GB (€2,300 for the complete machine).
Model sheet
ollama run llama3.3:70bWeights on Hugging Face: meta-llama/Llama-3.3-70B-Instruct
VRAM by precision and context
| Precision | 4k | 8k | 32k | 128k |
|---|---|---|---|---|
| FP16 | 150 GB | 151 GB | 159 GB | 188 GB |
| FP8 | 76 GB | 77 GB | 84 GB | 114 GB |
| Q8_0 | 80 GB | 81 GB | 89 GB | 119 GB |
| Q6_K | 62 GB | 64 GB | 71 GB | 101 GB |
| Q4_K_M | 46 GB | 47 GB | 55 GB | 85 GB |
| Q3_K_M | 38 GB | 39 GB | 47 GB | 76 GB |
One user. Each simultaneous user adds their own context cache.
Frequently asked questions
How much VRAM does Llama 3.3 70B need?
With an 8k-token context and one user, Llama 3.3 70B needs about 47 GB in Q4, 81 GB in Q8 and 151 GB in FP16.
What is the cheapest hardware to run Llama 3.3 70B?
In Q4, the cheapest option in our catalogue to run Llama 3.3 70B is Mini-PC Ryzen AI Max+ 395 128 GB, about €2,300 for the complete machine.
How fast does Llama 3.3 70B run?
Expect around 3.3 tokens per second for one user on Mini-PC Ryzen AI Max+ 395 128 GB in Q4 (estimate based on memory bandwidth).
On which hardware?
| Hardware | Cards | Speed | Machine |
|---|---|---|---|
| Mini-PC Ryzen AI Max+ 395 128 GB | 1 is enough | ~3.3 tok/s | €2,300 |
| NVIDIA DGX Spark 128 GB | 1 is enough | ~3.5 tok/s | €4,000 |
| Mac Studio M4 Max 128 GB | 1 is enough | ~7.1 tok/s | €4,400 |
| RTX 4060 Ti 16 GB | 4 cards | ~11 tok/s | €4,600 |
| Radeon RX 7800 XT 16 GB | 4 cards | ~24 tok/s | €4,800 |
| RTX 3090 24 GB (used) | 3 cards | ~27 tok/s | €5,050 |
| Radeon RX 7900 XTX 24 GB | 3 cards | ~28 tok/s | €5,650 |
| RTX 5090 32 GB | 2 cards | ~35 tok/s | €6,100 |
| RTX 4070 Ti SUPER 16 GB | 4 cards | ~26 tok/s | €6,200 |
| RTX 4080 SUPER 16 GB | 4 cards | ~29 tok/s | €7,000 |
| RTX 5080 16 GB | 4 cards | ~37 tok/s | €7,600 |
| RTX 6000 Ada 48 GB | 1 is enough | ~12 tok/s | €7,900 |
| RTX 4090 24 GB | 3 cards | ~29 tok/s | €8,800 |
| RTX PRO 6000 Blackwell 96 GB | 1 is enough | ~23 tok/s | €9,900 |
| Mac Studio M3 Ultra 512 GB | 1 is enough | ~11 tok/s | €11,500 |
| L40S 48 GB | 1 is enough | ~11 tok/s | €13,500 |
| A100 80 GB | 1 is enough | ~26 tok/s | €17,000 |
| H100 80 GB | 1 is enough | ~43 tok/s | €34,000 |
| H200 141 GB | 1 is enough | ~62 tok/s | €39,000 |
| B200 192 GB | 1 is enough | ~104 tok/s | €48,000 |
| RTX 3060 12 GB | doesn’t fit | — | — |
| RTX 4070 12 GB | doesn’t fit | — | — |
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: 2 × RTX 4060 Ti 16 GB · €1,800
From: Mini-PC Ryzen AI Max+ 395 128 GB · €2,300
From: Mac Studio M3 Ultra 512 GB · €11,500