Best home rig for Llama 3.2 3B

Older model. For current picks see 2025–2026 models.

Real, measured generation speeds for Llama 3.2 3B on home hardware, fastest first. Different quants/engines aren't directly comparable.

Build	Quant	Engine	tok/s
Intel Core Ultra 265K + RTX 4090		llama.cpp	334.67	source
Intel i9-13900K + RTX 5090		llama.cpp	271.4	source
Intel Core Ultra 265K + RTX 3080 Ti		llama.cpp	263.24	source
Ryzen 9 7900X + RTX 4090		llama.cpp	237.05	source
Ryzen 9 7950X + RTX 4080		llama.cpp	204.45	source
Ryzen 9 7950X + RTX 4070 Ti Super		llama.cpp	198.95	source
Intel Core Ultra 265K + RTX 4070 Ti		llama.cpp	196.15	source
Intel Core Ultra 265K + Radeon AI Pro R9700		llama.cpp	195.91	source
Intel Core Ultra 265K + RTX 2080 Ti		llama.cpp	180.38	source
Mac Studio M3 Ultra 512GB (Geerling test)		llama.cpp	154.6	source
Intel Core Ultra 265K + RTX A4000		llama.cpp	149.44	source
Ryzen 9 9950X + Radeon RX 7900 XT		llama.cpp	131.2	source
Mac Studio M1 Ultra (64-core GPU) 128GB		llama.cpp	127.6	source
Intel Core Ultra 265K + RTX 3060		llama.cpp	122.85	source
Dell Pro Max with GB10 (Nvidia DGX Spark)		llama.cpp	97.93	source
Framework Desktop (AMD Strix Halo) mainboard	Q4_K_M	llama.cpp Vulkan (iGPU, llama-bench tg128)	88.14	source
M4 Mac mini (10-core) 32GB		llama.cpp	41.31	source
AmpereOne A192-32X 512GB (CPU only)		llama.cpp	23.52	source

Rent before you buy: try this model on a cloud GPU for a few dollars first: Runpod (referral link; sign up with Google to get the new-user bonus)

Compare: DGX Spark vs Mac Studio vs Strix Halo: measured tok/s and tok/s per watt

FAQ

RTX 4090: 334.67 tok/s generation (llama.cpp), from https://github.com/geerlingguy/ai-benchmarks/issues/43.