Which VPS can run this model?
Prices checked
33 of 33 open models run on at least one of 527 plans without a GPU, checked .
Up to 4B
7–9B
12–15B
DeepSeek R1 · reasoning
DeepSeek-R1 Distill Qwen 14B
12 GB of RAM · from $9.04 a month · 187 plans fit
24–32B
DeepSeek R1 · reasoning
DeepSeek-R1 Distill Qwen 32B
23.4 GB of RAM · from $26.13 a month · 93 plans fit
70B+
DeepSeek R1 · reasoning
DeepSeek-R1 Distill Llama 70B
46.5 GB of RAM · from $34.18 a month · 34 plans fit
Mixture-of-experts
Meta Llama · chat · vision
Llama 4 Scout (109B MoE)
68.4 GB of RAM · from $65.69 a month · 9 plans fit
Embeddings & speech
Embeddings & speech · embedding
nomic-embed-text v1.5
2.1 GB of RAM · from $5.02 a month · 308 plans fit
How we count
Start from the model, not the server. For each open-weight model we work out the RAM a CPU-only VPS needs at Q4_K_M, pick the cheapest plans that fit and include a disk, estimate tokens per second from memory bandwidth, and price the local hardware alternative.
RAM needed is the weights at Q4_K_M (4.85 bits per weight times the parameters), plus the KV cache for a 8,192-token context (from each model's layer and head geometry), plus 1.5 GB for the OS and runtime. A plan is comfortable at 20% headroom, tight at exactly enough.
CPU generation is limited by memory bandwidth: every token streams the active weights once. We assume 4 GB/s per shared vCPU (at most 40) and 6 GB/s per dedicated vCPU (at most 60), divide by the active weight bytes, and show a band. Treat it as an order of magnitude, not a benchmark.
Model facts are fetched weekly from Hugging Face: parameter counts and licenses from the model API, exact GGUF file sizes from the quant repo and KV-cache geometry from the GGUF header (last refresh 26 Oct 2026). Only the one-line descriptions are written by hand. Prompt processing, batching and GPU offload are out of scope, so this is the CPU-only floor.