VPS Arena

Which VPS can run this model?

Start your search from the model, not the server. For 30 open-weight models we compute the RAM a CPU-only VPS needs at Q4_K_M, pick the cheapest plans that fit from all 280 tracked offers, estimate tokens per second from memory bandwidth, and price the local-hardware alternative — September 2026 data, refreshed daily.

Up to 4B

7–9B

12–15B

24–32B

70B+

Mixture-of-experts

Embeddings & speech

How these numbers are computed

RAM needed

Weights at Q4_K_M (4.85 bits per weight × parameters) + KV cache for a 8,192-token context (from each model's layer/head geometry) + 1.5 GB for the OS and runtime. A plan is “comfortable” at 20% headroom, “tight” at exactly enough.

Tokens per second

CPU generation is memory-bandwidth-bound: every token streams the active weights once. We assume 4 GB/s per shared vCPU (max 40) and 6 GB/s per dedicated vCPU (max 60), divide by active weight bytes, and show a ±20–40% band. Treat it as an order of magnitude, not a benchmark.

Catalog

Model facts are re-fetched weekly from Hugging Face — parameter counts and licenses from the model API, exact GGUF file sizes from the quant repo, KV-cache geometry from the GGUF header (last refresh 3 Sept 2026). Only the one-line descriptions are hand-written. Prompt processing, batching and GPU offload are out of scope — this is the honest CPU-only floor.