dexiio

What hardware runs Llama 3.1 8B?

Llama 3.1 8B is a 8.03B-parameter dense model (llama3.1 license). Aging, but the most finetuned model in history. The safe default when tooling compatibility matters more than raw smarts.

All figures assume an f16 KV cache, a 0.6 GB display reserve on the GPU, and 64 GB of DDR5 system RAM for the offload tiers. Tune these in the calculator.

Minimum VRAM by quant and context

QuantFile size8K16K32K64K128K
Q2_K *3.0 GB5.8 GB6.9 GB9 GB13.3 GB21.8 GB
Q3_K_M *3.7 GB6.6 GB7.7 GB9.8 GB14 GB22.5 GB
Q4_K_M4.6 GB7.4 GB8.5 GB10.6 GB14.9 GB23.4 GB
Q5_K_M5.3 GB8.2 GB9.3 GB11.4 GB15.6 GB24.1 GB
Q6_K6.1 GB9 GB10.1 GB12.2 GB16.4 GB24.9 GB
Q8_08.0 GB10.8 GB11.9 GB14 GB18.3 GB26.8 GB
IQ4_XS4.1 GB7 GB8.1 GB10.2 GB14.4 GB22.9 GB

Full-GPU figures: weights + f16 KV cache + overhead. * below our recommended floor of Q4_K_M.

GPU compatibility

GPU8K16K32K64K128K
NVIDIA GeForce RTX 5090Q8_0Q8_0Q8_0Q8_0Q8_0
NVIDIA GeForce RTX 5080Q8_0Q8_0Q8_0Q5_K_MQ8_0
NVIDIA GeForce RTX 5070 TiQ8_0Q8_0Q8_0Q5_K_MQ8_0
NVIDIA GeForce RTX 5070Q8_0Q8_0Q5_K_MIQ4_XSQ8_0
NVIDIA GeForce RTX 5060 Ti 16GBQ8_0Q8_0Q8_0Q5_K_MQ8_0
NVIDIA GeForce RTX 5060Q4_K_MIQ4_XSIQ4_XSQ8_0Q8_0
NVIDIA GeForce RTX 4090Q8_0Q8_0Q8_0Q8_0Q4_K_M
NVIDIA GeForce RTX 4080 SUPERQ8_0Q8_0Q8_0Q5_K_MQ8_0
NVIDIA GeForce RTX 4070 Ti SUPERQ8_0Q8_0Q8_0Q5_K_MQ8_0
NVIDIA GeForce RTX 4070Q8_0Q8_0Q5_K_MIQ4_XSQ8_0
NVIDIA GeForce RTX 4060 Ti 16GBQ8_0Q8_0Q8_0Q5_K_MQ8_0
NVIDIA GeForce RTX 4060Q4_K_MIQ4_XSIQ4_XSQ8_0Q8_0
NVIDIA GeForce RTX 3090Q8_0Q8_0Q8_0Q8_0Q4_K_M
NVIDIA GeForce RTX 3080 10GBQ6_KQ5_K_MIQ4_XSQ8_0Q8_0
NVIDIA GeForce RTX 3070Q4_K_MIQ4_XSIQ4_XSQ8_0Q8_0
NVIDIA GeForce RTX 3060 TiQ4_K_MIQ4_XSIQ4_XSQ8_0Q8_0
NVIDIA GeForce RTX 3060 12GBQ8_0Q8_0Q5_K_MIQ4_XSQ8_0
NVIDIA GeForce RTX 2080 TiQ8_0Q6_KQ4_K_MQ4_K_MQ8_0
NVIDIA GeForce GTX 1080 TiQ8_0Q6_KQ4_K_MQ4_K_MQ8_0
AMD Radeon RX 9070 XTQ8_0Q8_0Q8_0Q5_K_MQ8_0
AMD Radeon RX 7900 XTXQ8_0Q8_0Q8_0Q8_0Q4_K_M
AMD Radeon RX 7900 XTQ8_0Q8_0Q8_0Q8_0IQ4_XS
AMD Radeon RX 7800 XTQ8_0Q8_0Q8_0Q5_K_MQ8_0
AMD Radeon RX 7600 XTQ8_0Q8_0Q8_0Q5_K_MQ8_0
AMD Radeon RX 6800 XTQ8_0Q8_0Q8_0Q5_K_MQ8_0
AMD Radeon RX 6700 XTQ8_0Q8_0Q5_K_MIQ4_XSQ8_0
Intel Arc B580Q8_0Q8_0Q5_K_MIQ4_XSQ8_0
Intel Arc A770 16GBQ8_0Q8_0Q8_0Q5_K_MQ8_0
Apple M2 (16GB)Q8_0Q8_0Q5_K_MIQ4_XSQ8_0
Apple M2 Pro (32GB)Q8_0Q8_0Q8_0Q8_0Q4_K_M
Apple M2 Max (64GB)Q8_0Q8_0Q8_0Q8_0Q8_0
Apple M2 Ultra (128GB)Q8_0Q8_0Q8_0Q8_0Q8_0
Apple M3 (16GB)Q8_0Q8_0Q5_K_MIQ4_XSQ8_0
Apple M3 Pro (36GB)Q8_0Q8_0Q8_0Q8_0Q8_0
Apple M3 Max (64GB)Q8_0Q8_0Q8_0Q8_0Q8_0
Apple M3 Ultra (96GB)Q8_0Q8_0Q8_0Q8_0Q8_0
Apple M4 (16GB)Q8_0Q8_0Q5_K_MIQ4_XSQ8_0
Apple M4 Pro (48GB)Q8_0Q8_0Q8_0Q8_0Q8_0
Apple M4 Max (64GB)Q8_0Q8_0Q8_0Q8_0Q8_0
CPU only / integrated graphicsQ8_0Q8_0Q8_0Q8_0Q8_0

Fits on GPUExpert offloadPartial offloadCPU only

Quant guidance

Our floor for Llama 3.1 8B is Q4_K_M — below that, quality degrades faster than the VRAM savings are worth. Prefer the highest quant that still lands "Fits on GPU" at your context length in the table above.

Recommended run command

Q4_K_M at 32K on a NVIDIA GeForce RTX 2080 Ti-class GPU (Fits on GPU):

llama-server -m Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf -c 32768 --flash-attn -ngl 99

Frequently asked questions

How much VRAM does Llama 3.1 8B need?

At Q4_K_M and 32K context, Llama 3.1 8B needs about 10.6 GB of VRAM to run fully on GPU (weights + KV cache + overhead).

What is the smallest GPU that can run Llama 3.1 8B?

The NVIDIA GeForce RTX 2080 Ti (11 GB) is the smallest GPU in our set that runs Llama 3.1 8B well at 32K context.

What quantization should I use for Llama 3.1 8B?

We recommend Q4_K_M or higher. Q4_K_M weighs 4.6 GB (4.9 bits/weight); going below Q4_K_M costs noticeable quality on a model this size.

How long a context can Llama 3.1 8B handle?

Llama 3.1 8B supports up to 128K tokens. KV cache grows linearly with context.

Can I run Llama 3.1 8B without a GPU?

Yes, at reduced speed: on CPU with 64 GB of DDR5 it manages roughly 6 tokens/sec at 8K context (Q8_0).

Related guides