Clipper 16
$ 92 /month
- vCPU
- 8 × dedicated
- RAM
- 16 GB
- Storage
- 400 GB NVMe Gen4 SSD
- Transfer
- 30 TB
- IPv4 / IPv6
- 1 / /64 routed
Solution
VPS for AI workloads, LLM inference and RAG pipelines
Clipper 16 · $92/moClipper 32 · $174/mo
Short answer
CPU inference is viable for small models: an 8-billion-parameter model quantised to 4 bits runs at conversational speed on 8 dedicated cores with 16 GB of RAM. Anything larger wants a GPU. Most self-hosted AI work is actually the surrounding stack — vector database, embedding service, API gateway — and that runs perfectly well on a Clipper 32.
What you need
| Resource | What you actually need |
|---|---|
| CPU | 8+ dedicated cores with AVX-512 for CPU inference |
| RAM | 16 GB for an 8B model at Q4, 64 GB for 70B at Q4 |
| Disk | NVMe Gen4 — model weights are large and load-latency matters |
| GPU | Available in selected regions; ask before ordering |
Recommended plans
Clipper 32
$ 174 /month
- vCPU
- 12 × dedicated
- RAM
- 32 GB
- Storage
- 800 GB NVMe Gen4 SSD
- Transfer
- 40 TB
- IPv4 / IPv6
- 1 / /64 routed
Clipper 64
$ 329 /month
- vCPU
- 16 × dedicated
- RAM
- 64 GB
- Storage
- 1.6 TB NVMe Gen4 SSD
- Transfer
- 50 TB
- IPv4 / IPv6
- 1 / /64 routed
Recommended locations
Location is usually the decision that matters most for this workload — either because latency dominates, or because jurisdiction does.
Why OnionVPS for this
- High-frequency EPYC cores with AVX-512 make llama.cpp genuinely usable without a GPU.
- NVMe Gen4 loads a 40 GB model file in seconds rather than minutes.
- Self-hosting means no prompt ever leaves infrastructure you control — the actual reason most people do this.
- Privacy jurisdictions for teams processing sensitive documents through a model.
How to set it up
-
Deploy a Clipper 32 with Ubuntu 24.04
Twelve dedicated cores and 32 GB covers most CPU inference.
-
Install Ollama or llama.cpp
Ollama is the fastest path; llama.cpp gives finer control over quantisation.
-
Pull a quantised model
Q4_K_M is the usual quality-versus-speed sweet spot.
-
Put an API gateway in front
LiteLLM or a reverse proxy with authentication and rate limiting.
-
Add a vector database if you are building RAG
Qdrant or pgvector; both run comfortably alongside the model.
Frequently asked questions
Can you run an LLM on a VPS without a GPU?
Yes, for small and quantised models. An 8B model at Q4 produces roughly 8–15 tokens per second on 8 modern dedicated cores — fine for a personal assistant or a batch pipeline, too slow for a busy chat product.
How much RAM does a local LLM need?
Roughly the file size of the quantised weights plus 2 GB. An 8B model at Q4 is about 5 GB, a 70B model at Q4 about 40 GB.
Do you offer GPU instances?
In selected regions, yes. Contact us before ordering so we can confirm current availability and pricing.