The Only Rule That Matters
The model must fit entirely in RAM. If it doesn't, Ollama spills to disk and inference becomes unusably slow. We're talking minutes per response instead of seconds.
A rough sizing formula: model disk size in GB ≈ RAM needed. Add 2 GB headroom for the OS, which in practice means the 8 GB or 16 GB plan, not the entry one.
| VPS RAM | Max Model Size | Safe Choice |
|---|---|---|
| 4 GB | ~2 GB model | Llama 3.2 3B (Q4) |
| 8 GB | ~5 GB model | Llama 3.1 8B (Q4) |
| 16 GB | ~9 GB model | Qwen 2.5 14B (Q4) |
| 32 GB | ~20 GB model | DeepSeek R1 32B (Q4) |
Recommended Models by Use Case
General Chat and Summarization, Llama 3.1 8B
ollama pull llama3.1:8bThe most widely recommended starting model in 2026. Handles conversation, summarization, Q&A, and basic coding well. Requires 8 GB RAM. Inference on a 4-core CPU: about 4 tokens/second on shared cores, about 8 on dedicated ones.
Coding Assistant: Qwen 2.5 Coder 7B
ollama pull qwen2.5-coder:7bFine-tuned specifically for code. Outperforms Llama 3.1 8B on coding tasks at the same size. Requires 8 GB RAM.
Lightweight and Fast, Llama 3.2 3B
ollama pull llama3.2:3bFits in 4 GB RAM. Inference at 4–5 tokens/second on 2 shared vCPUs and 16–18 on 4 dedicated ones, enough for interactive chat with short prompts. Quality is noticeably lower than 7B models but fine for simple tasks.
Reasoning and Math, Phi-4 14B
ollama pull phi4:14bMicrosoft's Phi-4 punches above its weight on structured reasoning and math. Regularly outperforms larger 30B models on benchmarks. Requires 16 GB RAM. Not suitable for long documents, 16K context window.
Privacy-Focused Tasks: Mistral 7B
ollama pull mistral:7bStrong general performance, small footprint. A reliable alternative to Llama 3.1 for teams that prefer European-origin models.
Quantization: Q4 vs Q8
Ollama defaults to Q4_K_M quantization when you pull a model. This cuts memory by roughly two thirds compared to 16-bit weights with minimal quality loss, roughly the difference between a RAW photo and a high-quality JPEG.
For most use cases, Q4_K_M is the right choice. If you need higher quality and have extra RAM, pull with :q8_0:
ollama pull llama3.1:8b-instruct-q8_0What CPU Inference Is Good For
On a 4-core CPU, expect about 4 tokens per second for a 7B model on shared cores and about 8 on dedicated ones (our measurements). At the low end that is close to reading speed. This is:
- Fine for background processing, summarizing documents, extracting data, classifying text
- Usable for interactive chat if you're patient
- Not suitable for real-time code completion or latency-sensitive applications
For real-time interactive use, dedicated cores make the biggest difference: in our tests 4 dedicated cores (a newer CPU generation) were about twice as fast as 4 shared ones.
Running these on NoctHost
The plan follows the model you pick from the table above: the Standard plan (2 vCPU, 4 GB) runs 3B models, the Pro plan (4 vCPU, 8 GB) is the sweet spot for 7B and 8B, and the Beast plan (6 vCPU, 16 GB) handles 13B and 14B. For 32B models you need the Titan plan (8 vCPU, 32 GB). Billing is hourly from a prepaid balance you top up with crypto (no KYC) or a card, so you can size up for a heavy model and scale back down without a monthly commitment.