VPS Arena

Run Phi-4-mini 3.8B on a VPS.

September 2026 · 207 of 280 plans fit · re-ranked daily

MIT-licensed small model that is unusually good at math and function calling. On a CPU-only VPS it needs about 5 GB of RAM: 2.5 GB of Q4_K_M weights, 1 GB of KV cache for an 8,192-token context and 1.5 GB for the OS and runtime. 207 of the 280 plans in our index fit; the cheapest comfortable pick is Contabo's Cloud VPS 4 · 4 vCPU · 8 GB at $6.39/mo, streaming an estimated 4.1–8.2 tok/s.

RAM needed · CPU inference

5 GB

weights · Q4_K_M
2.5 GB
KV cache · 8K context
1 GB
OS + runtime headroom
1.5 GB

Other quants: Q8_0 weights 4.1 GB (near-lossless, about half of F16); F16 7.7 GB. Comfortable = 20% headroom over the total.

Model card

Size
3.84B parameters
Context
128K tokens
Kind
chat
Released
2025-02
License
MIT

Facts fetched from Hugging Face on 3 Sept 2026: exact GGUF file sizes, KV geometry from the GGUF header · 453,216 downloads.

ollama run phi4-mini:3.8bmodel card ↗Microsoft

Best VPS plans for Phi-4-mini 3.8B

ranked by estimated tokens/s per dollar, comfortable fits first

  1. 1
    ContaboCloud VPS 6 · 6 vCPU · 12 GB

    6 vCPU shared · 12 GB RAM · 200 GB SSD

    runs comfortably~6.2–12 tok/s$8.72/mo
  2. 2
    ContaboCloud VPS 4 · 4 vCPU · 8 GB

    4 vCPU shared · 8 GB RAM · 100 GB SSD

    runs comfortably~4.1–8.2 tok/s$6.39/mo
  3. 3
    netcupRS 2000 G12 · 8 dedicated cores · 16 GB

    8 vCPU dedicated · 16 GB RAM · 512 GB NVME

    runs comfortably~12–25 tok/s$20.93/mo
  4. 4
    ContaboCloud VPS 8 · 8 vCPU · 24 GB

    8 vCPU shared · 24 GB RAM · 300 GB SSD

    runs comfortably~8.2–17 tok/s$16.27/mo
  5. 5
    netcupRS 1000 G12 · 4 dedicated cores · 8 GB

    4 vCPU dedicated · 8 GB RAM · 256 GB NVME

    runs comfortably~6.2–12 tok/s$12.49/mo
  6. 6
    HetznerCX43 · 8 vCPU · 16 GB

    8 vCPU shared · 16 GB RAM · 160 GB NVME

    runs comfortably~8.2–17 tok/s$18.58/mo

Speed = effective memory bandwidth ÷ active weight bytes (4 GB/s per shared vCPU, 6 per dedicated), shown as a band. Real numbers depend on the host CPU generation, AVX-512/AMX support and how noisy the neighbours are — treat these as order-of-magnitude.

Or buy hardware · 14 reference machines fit

all machines →
  • Raspberry Pi 5 (16 GB)16 GB · 17 GB/s · ~1.9–2.9 tok/s$305= 48 mo of VPS
  • GeForce RTX 5060 Ti 16 GB (card only)16 GB · 448 GB/s · ~100+ tok/s$805*= 10+ yrs of VPS
  • Mac mini (M6, 16 GB)16 GB · 153 GB/s · ~35–53 tok/s$899= 10+ yrs of VPS
  • GeForce RTX 5070 Ti 16 GB (card only)16 GB · 896 GB/s · ~100+ tok/s$1,100*= 10+ yrs of VPS
  • Mac mini (M6, 32 GB)32 GB · 153 GB/s · ~35–53 tok/s$1,299= 10+ yrs of VPS
  • GeForce RTX 5080 16 GB (card only)16 GB · 960 GB/s · ~100+ tok/s$1,500*= 10+ yrs of VPS

Buy · per month

$8.78

$8.47 hardware + $0.31 power

Rent · per month

$6.39

Contabo Cloud VPS 4 · 4 vCPU · 8 GB

Break-even

The Raspberry Pi 5 (16 GB) pays for itself after 50 months of replacing the VPS — and it streams an estimated 1.9–2.9 tok/s against the VPS's CPU-only pace.

* approximate: August 2026 US retail median rather than list price. Local speed = peak bandwidth × 0.7 efficiency (0.35 on CPU-only boards) ÷ active weight bytes; unified-memory machines are assumed to give models 75% of their RAM. GPU cards need a host PC that is not included in the price.

Frequently asked

How much RAM does Phi-4-mini 3.8B need?
About 5 GB for CPU inference at Q4_K_M: 2.5 GB of weights, 1 GB of KV cache at 8,192 tokens of context, and 1.5 GB of headroom. Longer contexts need more KV cache (128 KB per token for this model).
What is the cheapest VPS that can run Phi-4-mini 3.8B?
Contabo Cloud VPS 4 · 4 vCPU · 8 GB (8 GB RAM, 4 vCPU) at $6.39/mo excl. VAT runs it comfortably as of 5 Sept 2026.
How fast will Phi-4-mini 3.8B run on a VPS without a GPU?
Roughly 6.2–12 tok/s on the top pick. Token generation is bound by memory bandwidth — each token streams the full weights once — so more and dedicated vCPUs help; a GPU or Apple-silicon machine is 10–50× faster.
Same lineupPhi-4 14BSame size classLlama 3.2 1BLlama 3.2 3BQwen3 0.6BQwen3 1.7BVPS for AI agents (API-hosted models) →