PerGB

Cheapest GPUs to run GLM-5.3 Flash 320B locally

GLM-5.3 Flash 320B needs about 195 GB of VRAM at Q4 with an 8K-token context (342 GB at Q8). A mixture of experts (18B active) with a compact attention cache. The cheapest way to hold it right now: 4 × A16 for $9,171, about 22–33 tokens/s.

The cheapest setups right now

Live on eBay
SetupVRAMSpeed (est.)Price nowPowerSoftware
4 × A16256 GB56.6 GB spare22–33 tokens/s$9,171the 4 cheapest listings1,000 WCUDA 11.1 and newer
4 × Instinct MI210256 GB56.6 GB spare32–54 tokens/s$15,249the 4 cheapest listings1,200 WROCm (on AMD's support list)
3 × RTX PRO 5000 Blackwell 72 GB216 GB17.6 GB spare37–55 tokens/s$27,744the 3 cheapest listings900 WCUDA 12.8 and newer
3 × A100 PCIe 80 GB240 GB41.6 GB spare53–80 tokens/s$39,498the 3 cheapest listings900 WCUDA 11.0 and newer
3 × RTX PRO 6000 Blackwell288 GB89.6 GB spare49–74 tokens/s$43,480the 3 cheapest listings1,800 Wover a 15 A circuitCUDA 12.8 and newer
2 × H200 NVL282 GB84.6 GB spare132–198 tokens/s$92,300the 2 cheapest listings1,200 WCUDA 11.8 and newer
3 × H100 NVL282 GB83.6 GB spare107–161 tokens/s$107,599the 3 cheapest listings1,200 WCUDA 11.8 and newer
3 × H100 PCIe240 GB41.6 GB spare55–82 tokens/s$108,112the 3 cheapest listings1,050 WCUDA 11.8 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.