TokenAssemble

Can an NVIDIA GeForce RTX 4090 run Llama-3.1-70B-Instruct?

Yes, with CPU offload. Llama-3.1-70B-Instruct needs 45.8GB at Q4_K_M with llama.cpp and 8K context; the NVIDIA GeForce RTX 4090 has 24GB of VRAM (23.5GB usable), so part of the model spills to system RAM — it loads, but expect the usable speed tier.

NVIDIA GeForce RTX 4090 · Llama-3.1-70B-Instruct · Q4_K_M

⚠️ Tight — grade C

UsablebetaCPU offload — speed varies heavily with layer split
weights 42.5 GBKV cache 2.68 GBoverhead 0.62 GBpool 23.5 GBneeds 45.8 GB — spills to system RAM

assumes llama.cpp · batch 1 · 8K context · fp16 KV cache · 64GB system RAM (CPU offload)

$1,599 MSRP · as of 2026-07-08

Affiliate link — we may earn a commission. Verdicts are computed before any link is attached.

Every quant, graded

NVIDIA GeForce RTX 4090 × Llama-3.1-70B-Instruct, llama.cpp, 8K context.

QuantFile sizeFitSpeed tier
Q8_075.0 GBWon't fit
Q6_K57.9 GBTight · Cpainful · offload · beta
Q5_K_M50.0 GBTight · Cpainful · offload · beta
Q4_K_M42.5 GBTight · Cusable · offload · beta

Llama-3.1-70B-Instruct on similar hardware

More models on the NVIDIA GeForce RTX 4090

Verdict computed from published specs and measured GGUF file sizes — see how the math works.