Llama · Meta

Llama 3.3 70B

See recommended builds Can my GPU run it?

Hardware requirement

Estimated

Running Llama 3.3 70B as Q4_K_M at ~8,192 tokens of context.

VRAM to run on GPU

45 GB

48 GB class card

System RAM

64 GB

recommended

Storage

2 TB

NVMe recommended

GPU importance

Critical

memory-bound workload

How this is calculated

Model weights derived	40.9 GB	70.554 billion parameters × 0.58 bytes/param (Q4_K_M) ≈ 40.9 GB.
KV cache (context memory) derived	2.7 GB	8,192 tokens × 80 layers × 2 × 8 KV heads × 128 head dim × 2 bytes ≈ 2.7 GB. Grows with context length.
Runtime overhead estimate	1.5 GB	Fixed VRAM overhead for activations and framework context.
Total VRAM to run fully on GPU estimate	45.1 GB	Weights + KV cache + overhead. This is what you need to keep the whole model on the graphics card for full speed.
Recommended system RAM estimate	64 GB	Headroom to load the model, run the OS, and hold any layers offloaded from the GPU. Rounded up to a standard kit size.
Storage estimate	2 TB NVMe recommended	Local model files are large (often tens of GB each). An NVMe SSD keeps load times reasonable.

Memory figures are estimates derived from published model facts using a fixed method. They are not benchmarks. We do not publish tokens-per-second figures because we have no measured source for them.

Quantizations

Smaller quantizations use less memory with some quality trade-off. The default below balances size and quality.

Variant	Bytes / param	Note
Q4_K_M default	0.58	Best size/quality balance — the usual pick.
Q5_K_M	0.69	A little larger, slightly better quality.
Q8_0	1.06	Near-lossless; roughly 8-bit weights.
FP16	2	Full half-precision weights; maximum fidelity, largest footprint.

Architecture facts sourced from the model creator's published configuration: https://huggingface.co/meta-llama.