Run Phi-4-mini 3.8B on a VPS.
Prices checked
The cheapest VPS that runs Phi-4-mini 3.8B is Leaseweb VPS 1 at $5.02 a month, with 6 GB of RAM for the 5 GB it needs, checked .
Best plans for Phi-4-mini 3.8B
Most tokens per second for your money, comfortable fits first.
- 1
VPS 1 · 4 vCPU / 6 GB · Europe
Leaseweb · 4C · 6 GB · ~4.1–8.2 tok/s$5.02/mo - 2
VPS 1 · 4 vCPU / 6 GB · UK
Leaseweb · 4C · 6 GB · ~4.1–8.2 tok/s$6.02/mo - 3
VPS 2 · 6 vCPU / 16 GB · Europe
Leaseweb · 6C · 16 GB · ~6.2–12 tok/s$9.04/mo - 4
VPS 1 · 4 vCPU / 6 GB · US
Leaseweb · 4C · 6 GB · ~4.1–8.2 tok/s$6.29/mo - 5
VPS 2 · 6 vCPU / 16 GB · UK
Leaseweb · 6C · 16 GB · ~6.2–12 tok/s$9.63/mo - 6
VPS 2 · 6 vCPU / 16 GB · US
Leaseweb · 6C · 16 GB · ~6.2–12 tok/s$9.89/mo
What it needs
RAM needed
5 GB
CPU inference at Q4_K_M
Weights
2.5 GB
Q4_K_M
KV cache
1 GB
8K context
OS and runtime
1.5 GB
headroom
MIT-licensed small model that is unusually good at math and function calling.
- Size
- 3.84B parameters
- Context
- 128K tokens
- Kind
- chat
- Released
- 2025-02
- License
- MIT
Facts from Hugging Face: exact GGUF file sizes, KV geometry from the GGUF header, 346,389 downloads.
Or rent a GPU
every GPU price1 provider rents the RTX A5000, with 24 GB of GPU memory against the 5 GB Phi-4-mini 3.8B needs.
Cheapest board that fits
RTX A5000
24 GB of GPU memory
On demand
$0.16per GPU an hour
RunPod
Held for a month
$116.80
730 hours of RTX A5000 · Community Cloud
Or buy hardware
all machines14 reference machines fit Phi-4-mini 3.8B.
Raspberry Pi 5 (16 GB)
Approximate16 GB · 17 GB/s · ~1.9–2.9 tok/s
$305= 61 mo of VPSGeForce RTX 5060 Ti 16 GB (card only)
Approximate16 GB · 448 GB/s · ~100+ tok/s
$456= 91 mo of VPSMac mini (M6, 16 GB)
16 GB · 153 GB/s · ~35–53 tok/s
$899= 10+ yrs of VPSGeForce RTX 5070 Ti 16 GB (card only)
Approximate16 GB · 896 GB/s · ~100+ tok/s
$1,190= 10+ yrs of VPSMac mini (M6, 32 GB)
32 GB · 153 GB/s · ~35–53 tok/s
$1,299= 10+ yrs of VPSGeForce RTX 5080 16 GB (card only)
Approximate16 GB · 960 GB/s · ~100+ tok/s
$1,650= 10+ yrs of VPS
Buy · per month
$8.78
$8.47 hardware + $0.31 power
Rent · per month
$5.02
Leaseweb VPS 1 · 4 vCPU / 6 GB · Europe
Break-even
The Raspberry Pi 5 (16 GB) pays for itself after 65 months of replacing the VPS, and it streams an estimated 1.9–2.9 tok/s against the VPS's CPU-only pace.
How we count
On a CPU-only VPS Phi-4-mini 3.8B 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. 260 of the 527 plans in the index fit and include a disk. A plan runs it comfortably with 20% headroom over the total, and is a tight fit with exactly enough.
Other quants: Q8_0 weights take 4.1 GB (near-lossless, about half of F16); F16 takes 7.7 GB.
The plans are ranked by estimated tokens per second per dollar, comfortable fits first. Speed is effective memory bandwidth divided by 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 and AMX support and how noisy the neighbours are, so treat them as an order of magnitude.
A GPU rate is per GPU for an hour, and the monthly figure is 730 hours of the whole instance at it, tax excluded.
A machine tagged approximate is priced at the typical US street price in August 2026 rather than its maker's own price. Local speed is peak bandwidth times 0.7 efficiency (0.35 on CPU-only boards) divided by active weight bytes; unified-memory machines are assumed to give models 75% of their RAM. GPU cards need a host PC that is not 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?
- Leaseweb VPS 1 · 4 vCPU / 6 GB · Europe (6 GB RAM, 4 vCPU) at $5.02/mo excl. VAT runs it comfortably as of 31 Oct 2026.
- How fast will Phi-4-mini 3.8B run on a VPS without a GPU?
- Roughly 4.1–8.2 tok/s on the top pick. Token generation is limited by memory bandwidth, because each new token reads all of the weights once. More vCPUs help, dedicated ones most, but a GPU or an Apple silicon machine is 10 to 50 times faster.
Same lineup
Same size class