Cheapest GPUs to run Qwen3.8 Flash-Next 180B locally
Qwen3.8 Flash-Next 180B needs about 110 GB of VRAM at Q4 with an 8K-token context (192 GB at Q8). A mixture of experts: 6B active, plus a 51B n-gram embedding. The cheapest way to hold it right now: 4 × Instinct MI50 32 GB for $2,726, about 59–101 tokens/s.
The cheapest setups right now
Live on eBay| Setup | VRAM | Speed (est.) | Price now | Power | Software |
|---|---|---|---|---|---|
| 4 × Instinct MI50 32 GB | 128 GB14.2 GB spare | 59–101 tokens/s | $2,726the 4 cheapest listings | 1,200 W | Not on ROCm's support list (older ROCm or Vulkan) |
| 4 × Radeon Pro W6800 | 128 GB14.2 GB spare | 30–51 tokens/s | $3,540the 4 cheapest listings | 1,000 W | ROCm (on AMD's support list) |
| 4 × Tesla V100 PCIe 32 GB | 128 GB14.2 GB spare | 74–111 tokens/s | $3,909the 4 cheapest listings | 1,000 W | CUDA 9.0 to 12.9 (not CUDA 13) |
| 2 × A16 | 128 GB16.2 GB spare | 66–99 tokens/s | $4,439the 2 cheapest listings | 500 W | CUDA 11.1 and newer |
| 4 × Tesla V100S PCIe | 128 GB14.2 GB spare | 94–140 tokens/s | $4,606the 4 cheapest listings | 1,000 W | CUDA 9.0 to 12.9 (not CUDA 13) |
| 4 × Instinct MI100 | 128 GB14.2 GB spare | 71–122 tokens/s | $5,523the 4 cheapest listings | 1,200 W | ROCm (on AMD's support list) |
| 3 × Quadro RTX 8000 | 144 GB31.2 GB spare | 55–83 tokens/s | $6,000the 3 cheapest listings | 780 W | CUDA 10.0 and newer |
| 3 × Arc Pro B60 48 GB | 144 GB31.2 GB spare | 26–45 tokens/s | $6,159the 3 cheapest listings | 600 W | No CUDA (oneAPI / Vulkan) |
| 4 × Quadro GV100 | 128 GB14.2 GB spare | 72–108 tokens/s | $6,298the 4 cheapest listings | 1,000 W | CUDA 9.0 to 12.9 (not CUDA 13) |
| 4 × Radeon Pro W7800 32 GB | 128 GB14.2 GB spare | 33–57 tokens/s | $7,501the 4 cheapest listings | 1,040 W | ROCm (on AMD's support list) |
| 2 × Instinct MI210 | 128 GB16.2 GB spare | 95–162 tokens/s | $7,598the 2 cheapest listings | 600 W | ROCm (on AMD's support list) |
| 4 × Radeon AI PRO R9700 | 128 GB14.2 GB spare | 37–63 tokens/s | $8,938the 4 cheapest listings | 1,200 W | ROCm (on AMD's support list) |
| 3 × A40 | 144 GB31.2 GB spare | 57–86 tokens/s | $10,025the 3 cheapest listings | 900 W | CUDA 11.1 and newer |
| 4 × GeForce RTX 4080 32 GB | 128 GB14.2 GB spare | 59–89 tokens/s | $10,200the 4 cheapest listings | 1,280 Wover a 15 A circuit | CUDA 11.8 and newer |
| 4 × GeForce RTX 4080 Super 32 GB | 128 GB14.2 GB spare | 61–91 tokens/s | $10,700the 4 cheapest listings | 1,280 Wover a 15 A circuit | CUDA 11.8 and newer |
Per question: scaled from the model's reported speed on an RTX PRO 6000 by each card's compute and memory bandwidth, within about half either way. Licenses, scores and sources: decision models.
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.