Mixtral 8x7B VRAM Requirements

46.70B parameters, 32 layers, 8 key-value heads. At Q4_K_M with a 8K context it needs 26.3 GB — 13 of 30 accelerators here can hold it.

A mixture of experts: 8 experts per layer, 2 of them running per token. All of them have to be in memory, so the 46.7B above is what you need to hold — but only 12.88B is doing work at any moment, which is what sets the speed.

Your setup

About 22.1 GB of that is usable for a model.

The usual choice for running locally. Quality loss is small and it is what most GGUF downloads default to.

8K tokens max 32K

Longer context costs memory linearly — this model uses 128.0 KB per token.

Running Mixtral 8x7B

26.3 GB of memory needed

Will not fit on a GeForce RTX 4090

Weights
24.5 GB
KV cache
1.0 GB
Headroom
-4.2 GB
Max context
0
26.3 GB used 22.1 GB usable

Where the memory goes

Mixtral 8x7B memory breakdown
Component Memory
Model weights (Q4_K_M) 24.5 GB
KV cache at 8K tokens 1.0 GB
Runtime allowance 0.8 GB
Total 26.3 GB

Weights are only half the question

A 46.7B model at Q4_K_M is 24.5 GB of weights. That part is fixed — it is the file on disk, and it does not change while the model runs. What changes is the key-value cache, which grows with every token of context you give it.

Mixtral 8x7B spends 128.0 KB per token. At 8K that is 1.0 GB, comfortably smaller than the weights. At its full 32K it is 4.0 GB — still under the weights, which is unusual for a long-context model. That is the trap: the model loads fine and then fills the card once a conversation gets long.

The reason the cache is as small as it is comes down to grouped-query attention. Mixtral 8x7B has 32 query heads but only 8 key-value heads, and the cache is sized by the latter. Calculators that use the query head count overstate it 4-fold, which is where the wildly pessimistic numbers people quote usually come from. The name is not the size. Eight experts of 7B each come to 46.7B rather than 56B, because the attention layers and embeddings are shared and only the MLPs are replicated. Its config states no separate expert width, so the experts use the model's own intermediate size.

By quantisation

Mixtral 8x7B memory by quantisation
Format Bits/weight Weights Total at 8K
FP16 / BF16 16 87.0 GB 88.8 GB
Q8_0 8.5 46.2 GB 48.0 GB
Q6_K 6.6 35.9 GB 37.7 GB
Q5_K_M 5.5 29.9 GB 31.7 GB
Q4_K_M 4.5 24.5 GB 26.3 GB
Q3_K_M 3.9 21.2 GB 23.0 GB

By context length

Mixtral 8x7B memory by context length at Q4_K_M
Context KV cache Total
2K tokens 0.3 GB 25.5 GB
4K tokens 0.5 GB 25.8 GB
8K tokens 1.0 GB 26.3 GB
16K tokens 2.0 GB 27.3 GB
32K tokens 4.0 GB 29.3 GB

Will it run on your hardware?

Which accelerators run Mixtral 8x7B at Q4_K_M
Accelerator Memory Runs it Headroom Max context
GeForce RTX 5090 32 GB Yes 3.1 GB 32K
GeForce RTX 4090 24 GB No
GeForce RTX 3090 24 GB No
GeForce RTX 5080 16 GB No
GeForce RTX 4080 SUPER 16 GB No
GeForce RTX 5070 Ti 16 GB No
GeForce RTX 4070 Ti SUPER 16 GB No
GeForce RTX 5070 12 GB No
GeForce RTX 4070 12 GB No
GeForce RTX 3060 12GB 12 GB No
GeForce RTX 4060 Ti 16GB 16 GB No
GeForce RTX 3080 10 GB No
Radeon RX 7900 XTX 24 GB No
Radeon RX 7900 XT 20 GB No
Radeon RX 9070 XT 16 GB No
2× RTX 3090 (48 GB) 48 GB Yes 17.9 GB 32K
2× RTX 4090 (48 GB) 48 GB Yes 17.9 GB 32K
4× RTX 3090 (96 GB) 96 GB Yes 62.0 GB 32K
Mac (M4 Max, 64 GB) 64 GB Yes 21.7 GB 32K
Mac (M4 Max, 128 GB) 128 GB Yes 69.7 GB 32K
Mac (M4 Pro, 48 GB) 48 GB Yes 9.7 GB 32K
Mac (M4, 24 GB) 24 GB No
Mac (M3 Max, 128 GB) 128 GB Yes 69.7 GB 32K
Mac Studio (M2 Ultra, 192 GB) 192 GB Yes 117.7 GB 32K
Mac (M1 Max, 32 GB) 32 GB No
Mac (M2, 16 GB) 16 GB No
NVIDIA A100 40GB 40 GB Yes 11.7 GB 32K
NVIDIA A100 80GB 80 GB Yes 49.7 GB 32K
NVIDIA H100 80GB 80 GB Yes 49.7 GB 32K
NVIDIA L40S 48GB 48 GB Yes 19.3 GB 32K

At Q4_K_M with a 8K context. "Only just" means the model uses more than 95% of usable memory — it will load and then fall over as soon as anything else touches the device.

Frequently asked questions

How much VRAM does Mixtral 8x7B need?
About 26.3 GB at Q4_K_M with a 8K context — 24.5 GB of weights, 1.0 GB of key-value cache and a 0.8 GB runtime allowance. Unquantised at FP16 it needs 88.8 GB, which is why almost nobody runs it that way locally.
What GPU do I need to run Mixtral 8x7B?
The smallest card here that runs it is the GeForce RTX 5090 with 32 GB. 13 of the 30 accelerators on this page can hold it at the default quantisation. Apple Silicon deserves a mention: unified memory means a Mac can hold models no consumer graphics card can, though it runs them more slowly.
Does context length change how much memory Mixtral 8x7B needs?
Yes, and linearly. This model spends 128.0 KB per token of context, so 8K costs 1.0 GB and its full 32K costs 4.0 GB. On a model with a long context window the cache can end up larger than the quantised weights, which catches people out — the weights fit and then the conversation does not.
Why is the cache smaller than other calculators say?
Because Mixtral 8x7B uses grouped-query attention: 8 key-value heads shared across 32 query heads. The cache is sized by the key-value head count, not the query head count, so using the latter overstates it by a factor of 4. Plenty of calculators still get this wrong.
Which quantisation should I use?
Q4_K_M is the usual answer — the quality loss is small and it is what most GGUF downloads default to. Step up to Q5_K_M or Q6_K if you have memory to spare, since quality improves and the cost is modest. Q8_0 is effectively lossless at half the size of FP16. Drop to Q3_K_M only to fit a larger model that otherwise would not run at all; a bigger model at a rougher quantisation usually beats a smaller one at a fine quantisation.

Architecture: Hugging Face config.json. GGUF K-quant effective bit widths as produced by llama.cpp.

An estimate of memory, not a benchmark. It counts model weights, the key-value cache at the context you choose, and a fixed runtime allowance. Actual usage moves with the runtime, the batch size, whether flash attention is on, and how much the operating system has already taken from a shared memory pool. Treat a result within a gigabyte of your card's capacity as 'probably not' rather than 'just fits'.

Data on this page last verified .

    to move to open