PerGB

Cheapest GPUs to run Llama 4 Scout 109B locally

Llama 4 Scout 109B needs about 68 GB of VRAM at Q4 with an 8K-token context (118 GB at Q8). A mixture of experts: 17B active. The cheapest way to hold it right now: 3 × Tesla P40 for $810.00, about 10–15 tokens/s.

The cheapest setups right now

Live on eBay
SetupVRAMSpeed (est.)Price nowPowerSoftware
3 × Tesla P4072 GB0.8 GB spare10–15 tokens/s$810.00the 3 cheapest listings750 WCUDA 8.0 to 12.9 (not CUDA 13)
3 × Instinct MI50 32 GB96 GB24.8 GB spare21–36 tokens/s$2,027the 3 cheapest listings900 WNot on ROCm's support list (older ROCm or Vulkan)
3 × Arc Pro B60 24 GB72 GB0.8 GB spare9.3–16 tokens/s$2,549the 3 cheapest listings600 WNo CUDA (oneAPI / Vulkan)
3 × Radeon Pro W680096 GB24.8 GB spare10–18 tokens/s$2,650the 3 cheapest listings750 WROCm (on AMD's support list)
4 × GeForce RTX 2080 Ti 22 GB88 GB15.8 GB spare18–27 tokens/s$2,679the 4 cheapest listings1,000 WCUDA 10.0 and newer
3 × Titan RTX72 GB0.8 GB spare20–29 tokens/s$2,840the 3 cheapest listings840 WCUDA 10.0 and newer
3 × Tesla V100 PCIe 32 GB96 GB24.8 GB spare26–39 tokens/s$2,929the 3 cheapest listings750 WCUDA 9.0 to 12.9 (not CUDA 13)
3 × Quadro RTX 600072 GB0.8 GB spare20–29 tokens/s$2,970the 3 cheapest listings780 WCUDA 10.0 and newer
4 × GeForce RTX 3080 20 GB80 GB7.8 GB spare27–40 tokens/s$3,364the 4 cheapest listings1,280 Wover a 15 A circuitCUDA 11.1 and newer
3 × Tesla V100S PCIe96 GB24.8 GB spare33–50 tokens/s$3,436the 3 cheapest listings750 WCUDA 9.0 to 12.9 (not CUDA 13)
4 × Radeon RX 7900 XT80 GB7.8 GB spare16–28 tokens/s$3,489the 4 cheapest listings1,260 Wover a 15 A circuitROCm (on AMD's support list)
3 × Radeon RX 7900 XTX72 GB0.8 GB spare20–34 tokens/s$3,550the 3 cheapest listings1,065 WROCm (on AMD's support list)
2 × Quadro RTX 800096 GB25.8 GB spare20–29 tokens/s$4,000the 2 cheapest listings520 WCUDA 10.0 and newer
3 × Instinct MI10096 GB24.8 GB spare25–43 tokens/s$4,029the 3 cheapest listings900 WROCm (on AMD's support list)
2 × Arc Pro B60 48 GB96 GB25.8 GB spare9.3–16 tokens/s$4,104the 2 cheapest listings400 WNo CUDA (oneAPI / Vulkan)

A set of cards is priced at that many of the cheapest live listings, shipping included. Speed: an estimate from memory bandwidth for one user chatting, within a third either way.

How it works

The memory a model needs is its weights at the quantisation you pick, plus the attention cache for the context and a little for working buffers. Q4 is what most people run at home, close to full quality at a third of the size. Each card keeps about 1 GB for itself, so two 24 GB cards hold 46 GB.

Writing each token reads every active weight once, so the speed is about the memory bandwidth divided by the model's size, times what the software gets out of the card (less off CUDA). Cards sharing a model layer by layer take turns, so a set runs about as fast as one card would. Treat it as a rough figure for one person chatting; a long prompt takes longer to read in.

Power is the cards' rated power. Over about 1,440 W with the rest of the PC, a set needs more than one 15 A circuit (or a 240 V one), and older cards often need a power cable adapter, which each card's page mentions.