Llama 3.1 8B VRAM Requirements

8.03B parameters, 32 layers, 8 key-value heads. At Q4_K_M with a 8K context it needs 6.0 GB — 30 of 30 accelerators here can hold it.

Your GPU or Mac

About 22.1 GB of that is usable for a model.

Quantisation

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

Context length

8K tokens max 128K

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

KV cache precision

Running Llama 3.1 8B

6.0 GB of memory needed

Runs comfortably on a GeForce RTX 4090

Weights: 4.2 GB
KV cache: 1.0 GB
Headroom: 16.1 GB
Max context: 128K

6.0 GB used 22.1 GB usable

Where the memory goes

Llama 3.1 8B memory breakdown
Component	Memory
Model weights (Q4_K_M)	4.2 GB
KV cache at 8K tokens	1.0 GB
Runtime allowance	0.8 GB
Total	6.0 GB

Weights are only half the question

A 8.0B model at Q4_K_M is 4.2 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.

Llama 3.1 8B spends 128.0 KB per token. At 8K that is 1.0 GB, comfortably smaller than the weights. At its full 128K it is 16.0 GB — larger than the weights themselves. 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. Llama 3.1 8B 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 128k context is the headline, but it is also the reason a 24 GB card fills up: the KV cache at full context is larger than the quantised weights.

By quantisation

Llama 3.1 8B memory by quantisation
Format	Bits/weight	Weights	Total at 8K
FP16 / BF16	16	15.0 GB	16.8 GB
Q8_0	8.5	7.9 GB	9.7 GB
Q6_K	6.6	6.2 GB	8.0 GB
Q5_K_M	5.5	5.1 GB	6.9 GB
Q4_K_M	4.5	4.2 GB	6.0 GB
Q3_K_M	3.9	3.6 GB	5.4 GB

By context length

Llama 3.1 8B memory by context length at Q4_K_M
Context	KV cache	Total
2K tokens	0.3 GB	5.3 GB
4K tokens	0.5 GB	5.5 GB
8K tokens	1.0 GB	6.0 GB
16K tokens	2.0 GB	7.0 GB
32K tokens	4.0 GB	9.0 GB
64K tokens	8.0 GB	13.0 GB
128K tokens	16.0 GB	21.0 GB

Will it run on your hardware?

Which accelerators run Llama 3.1 8B at Q4_K_M
Accelerator	Memory	Runs it	Headroom	Max context
GeForce RTX 5090	32 GB	Yes	23.4 GB	128K
GeForce RTX 4090	24 GB	Yes	16.1 GB	128K
GeForce RTX 3090	24 GB	Yes	16.1 GB	128K
GeForce RTX 5080	16 GB	Yes	8.7 GB	78K
GeForce RTX 4080 SUPER	16 GB	Yes	8.7 GB	78K
GeForce RTX 5070 Ti	16 GB	Yes	8.7 GB	78K
GeForce RTX 4070 Ti SUPER	16 GB	Yes	8.7 GB	78K
GeForce RTX 5070	12 GB	Yes	5.0 GB	48K
GeForce RTX 4070	12 GB	Yes	5.0 GB	48K
GeForce RTX 3060 12GB	12 GB	Yes	5.0 GB	48K
GeForce RTX 4060 Ti 16GB	16 GB	Yes	8.7 GB	78K
GeForce RTX 3080	10 GB	Yes	3.2 GB	34K
Radeon RX 7900 XTX	24 GB	Yes	16.1 GB	128K
Radeon RX 7900 XT	20 GB	Yes	12.4 GB	107K
Radeon RX 9070 XT	16 GB	Yes	8.7 GB	78K
2× RTX 3090 (48 GB)	48 GB	Yes	38.2 GB	128K
2× RTX 4090 (48 GB)	48 GB	Yes	38.2 GB	128K
4× RTX 3090 (96 GB)	96 GB	Yes	82.3 GB	128K
Mac (M4 Max, 64 GB)	64 GB	Yes	42.0 GB	128K
Mac (M4 Max, 128 GB)	128 GB	Yes	90.0 GB	128K
Mac (M4 Pro, 48 GB)	48 GB	Yes	30.0 GB	128K
Mac (M4, 24 GB)	24 GB	Yes	10.1 GB	89K
Mac (M3 Max, 128 GB)	128 GB	Yes	90.0 GB	128K
Mac Studio (M2 Ultra, 192 GB)	192 GB	Yes	138.0 GB	128K
Mac (M1 Max, 32 GB)	32 GB	Yes	15.4 GB	128K
Mac (M2, 16 GB)	16 GB	Yes	4.7 GB	46K
NVIDIA A100 40GB	40 GB	Yes	32.0 GB	128K
NVIDIA A100 80GB	80 GB	Yes	70.0 GB	128K
NVIDIA H100 80GB	80 GB	Yes	70.0 GB	128K
NVIDIA L40S 48GB	48 GB	Yes	39.6 GB	128K

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 Llama 3.1 8B need?

About 6.0 GB at Q4_K_M with a 8K context — 4.2 GB of weights, 1.0 GB of key-value cache and a 0.8 GB runtime allowance. Unquantised at FP16 it needs 16.8 GB, which is why almost nobody runs it that way locally.

What GPU do I need to run Llama 3.1 8B?

The smallest card here that runs it is the GeForce RTX 3080 with 10 GB. 30 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 Llama 3.1 8B needs?

Yes, and linearly. This model spends 128.0 KB per token of context, so 8K costs 1.0 GB and its full 128K costs 16.0 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 Llama 3.1 8B 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 (public mirror of the gated original). 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 2 September 2026.