PerGB

Cheapest GPUs to run Qwen3.8 27B locally

Qwen3.8 27B needs about 18 GB of VRAM at Q4 with an 8K-token context (31 GB at Q8). The cheapest way to hold it right now: 2 × Tesla P100 PCIe 12 GB for $257.00, about 16–24 tokens/s.

The cheapest setups right now

Live on eBay
SetupVRAMSpeed (est.)Price nowPowerSoftware
2 × Tesla P100 PCIe 12 GB24 GB4.1 GB spare16–24 tokens/s$257.00the 2 cheapest listings500 WCUDA 8.0 to 12.9 (not CUDA 13)
Tesla P4024 GB5.1 GB spare10–15 tokens/s$270.00250 WCUDA 8.0 to 12.9 (not CUDA 13)
2 × Tesla P100 PCIe 16 GB32 GB12.1 GB spare22–33 tokens/s$345.00the 2 cheapest listings500 WCUDA 8.0 to 12.9 (not CUDA 13)
2 × Radeon VII32 GB12.1 GB spare21–36 tokens/s$503.00the 2 cheapest listings600 WNot on ROCm's support list (older ROCm or Vulkan)
2 × Instinct MI50 16 GB32 GB12.1 GB spare21–36 tokens/s$538.62the 2 cheapest listings600 WNot on ROCm's support list (older ROCm or Vulkan)
3 × GeForce RTX 2070 Super24 GB3.1 GB spare13–20 tokens/s$539.98the 3 cheapest listings645 WCUDA 10.0 and newer
2 × GeForce RTX 2080 Ti 11 GB22 GB2.1 GB spare18–27 tokens/s$540.00the 2 cheapest listings500 WCUDA 10.0 and newer
2 × Tesla V100 PCIe 12 GB24 GB4.1 GB spare27–40 tokens/s$543.85the 2 cheapest listings500 WCUDA 9.0 to 12.9 (not CUDA 13)
3 × Quadro RTX 400024 GB3.1 GB spare12–19 tokens/s$549.97the 3 cheapest listings480 WCUDA 10.0 and newer
3 × GeForce RTX 207024 GB3.1 GB spare13–20 tokens/s$555.49the 3 cheapest listings525 WCUDA 10.0 and newer
2 × Radeon RX 6700 XT24 GB4.1 GB spare8.0–14 tokens/s$580.00the 2 cheapest listings460 WNot on ROCm's support list (older ROCm or Vulkan)
3 × GeForce RTX 208024 GB3.1 GB spare13–20 tokens/s$585.00the 3 cheapest listings645 WCUDA 10.0 and newer
3 × GeForce RTX 2060 Super24 GB3.1 GB spare13–20 tokens/s$585.51the 3 cheapest listings525 WCUDA 10.0 and newer
2 × GeForce RTX 2060 12 GB24 GB4.1 GB spare10.0–15 tokens/s$589.99the 2 cheapest listings320 WCUDA 10.0 and newer
GeForce RTX 3080 20 GB20 GB1.1 GB spare27–41 tokens/s$597.36320 WCUDA 11.1 and newer

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.