Mixtral 8x7B: how much VRAM?
In Q4_K_M with 8k tokens of context, Mixtral 8x7B needs about 31 GB of memory. The cheapest option: 2 × RTX 4060 Ti 16 GB (€1,800 for the complete machine).
Model sheet
ollama run mixtral:8x7bWeights on Hugging Face: mistralai/Mixtral-8x7B-Instruct-v0.1
VRAM by precision and context
| Precision | 4k | 8k | 32k | 128k |
|---|---|---|---|---|
| FP16 | 99 GB | 99 GB | 102 GB | 114 GB |
| FP8 | 50 GB | 50 GB | 53 GB | 65 GB |
| Q8_0 | 53 GB | 53 GB | 56 GB | 68 GB |
| Q6_K | 41 GB | 42 GB | 45 GB | 57 GB |
| Q4_K_M | 30 GB | 31 GB | 34 GB | 46 GB |
| Q3_K_M | 25 GB | 25 GB | 28 GB | 40 GB |
One user. Each simultaneous user adds their own context cache.
Frequently asked questions
How much VRAM does Mixtral 8x7B need?
With an 8k-token context and one user, Mixtral 8x7B needs about 31 GB in Q4, 53 GB in Q8 and 99 GB in FP16.
What is the cheapest hardware to run Mixtral 8x7B?
In Q4, the cheapest option in our catalogue to run Mixtral 8x7B is 2 × RTX 4060 Ti 16 GB, about €1,800 for the complete machine.
How fast does Mixtral 8x7B run?
Expect around 30 tokens per second for one user on 2 × RTX 4060 Ti 16 GB in Q4 (estimate based on memory bandwidth).
On which hardware?
| Hardware | Cards | Speed | Machine |
|---|---|---|---|
| RTX 4060 Ti 16 GB | 2 cards | ~30 tok/s | €1,800 |
| Radeon RX 7800 XT 16 GB | 2 cards | ~65 tok/s | €1,900 |
| Mini-PC Ryzen AI Max+ 395 128 GB | 1 is enough | ~18 tok/s | €2,300 |
| RTX 3090 24 GB (used) | 2 cards | ~98 tok/s | €2,400 |
| RTX 4070 Ti SUPER 16 GB | 2 cards | ~70 tok/s | €2,600 |
| Radeon RX 7900 XTX 24 GB | 2 cards | ~100 tok/s | €2,800 |
| RTX 4080 SUPER 16 GB | 2 cards | ~77 tok/s | €3,000 |
| RTX 5080 16 GB | 2 cards | ~100 tok/s | €3,300 |
| RTX 5090 32 GB | 1 is enough | ~125 tok/s | €3,500 |
| RTX 3060 12 GB | 3 cards | ~56 tok/s | €3,640 |
| NVIDIA DGX Spark 128 GB | 1 is enough | ~19 tok/s | €4,000 |
| Mac Studio M4 Max 128 GB | 1 is enough | ~38 tok/s | €4,400 |
| RTX 4070 12 GB | 3 cards | ~79 tok/s | €4,450 |
| RTX 4090 24 GB | 2 cards | ~105 tok/s | €4,900 |
| RTX 6000 Ada 48 GB | 1 is enough | ~67 tok/s | €7,900 |
| RTX PRO 6000 Blackwell 96 GB | 1 is enough | ~125 tok/s | €9,900 |
| Mac Studio M3 Ultra 512 GB | 1 is enough | ~57 tok/s | €11,500 |
| L40S 48 GB | 1 is enough | ~60 tok/s | €13,500 |
| A100 80 GB | 1 is enough | ~142 tok/s | €17,000 |
| H100 80 GB | 1 is enough | ~234 tok/s | €34,000 |
| H200 141 GB | 1 is enough | ~335 tok/s | €39,000 |
| B200 192 GB | 1 is enough | ~558 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: 2 × RTX 3060 12 GB · €1,460
From: 2 × RTX 3060 12 GB · €1,460
From: 2 × RTX 3060 12 GB · €1,460