VPS Arena

Run Qwen3 30B-A3B (MoE) on a VPS.

September 2026 · 68 of 280 plans fit · re-ranked daily

The CPU darling: 30B of knowledge, 3B active per token, so it streams like a small model on a 24–32 GB VPS. On a CPU-only VPS it needs about 20.9 GB of RAM: 18.6 GB of Q4_K_M weights, 819 MB of KV cache for an 8,192-token context and 1.5 GB for the OS and runtime. 68 of the 280 plans in our index fit; the cheapest comfortable pick is Contabo's Cloud VPS 12 · 12 vCPU · 48 GB at $29.05/mo, streaming an estimated 6–12 tok/s.

RAM needed · CPU inference

20.9 GB

weights · Q4_K_M
18.6 GB
KV cache · 8K context
819 MB
OS + runtime headroom
1.5 GB

Other quants: Q8_0 weights 32.5 GB (near-lossless, about half of F16); F16 61.1 GB. Comfortable = 20% headroom over the total.

Model card

Size
31B total · 3.3B active
Context
40K tokens
Kind
chat
Released
2025-04
License
Apache-2.0

Facts fetched from Hugging Face on 3 Sept 2026: exact GGUF file sizes, KV geometry from the GGUF header · 2,375,624 downloads.

ollama run qwen3:30b-a3bmodel card ↗Alibaba

Best VPS plans for Qwen3 30B-A3B (MoE)

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

  1. 1
    netcupRS 4000 G12 · 12 dedicated cores · 32 GB

    12 vCPU dedicated · 32 GB RAM · 1024 GB NVME

    runs comfortably~9–18 tok/s$38.99/mo
  2. 2
    ContaboCloud VPS 12 · 12 vCPU · 48 GB

    12 vCPU shared · 48 GB RAM · 400 GB SSD

    runs comfortably~6–12 tok/s$29.05/mo
  3. 3
    netcupVPS 4000 G12 · 12 vCore · 32 GB

    12 vCPU shared · 32 GB RAM · 1024 GB NVME

    runs comfortably~6–12 tok/s$31.66/mo
  4. 4
    HetznerCX53 · 16 vCPU · 32 GB

    16 vCPU shared · 32 GB RAM · 320 GB NVME

    runs comfortably~6–12 tok/s$34.27/mo
  5. 5
    ContaboCloud VPS 16 · 16 vCPU · 64 GB

    16 vCPU shared · 64 GB RAM · 500 GB SSD

    runs comfortably~6–12 tok/s$43/mo
  6. 6
    netcupRS 8000 G12 · 16 dedicated cores · 64 GB

    16 vCPU dedicated · 64 GB RAM · 2048 GB NVME

    runs comfortably~9–18 tok/s$69.70/mo

Speed = effective memory bandwidth ÷ active weight bytes (4 GB/s per shared vCPU, 6 per dedicated, ×0.5 for MoE routing overhead), 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 · 8 reference machines fit

all machines →
  • Mac mini (M6, 32 GB)32 GB · 153 GB/s · ~20–31 tok/s$1,299= 45 mo of VPS
  • Framework Desktop (Ryzen AI Max+ 395, 64 GB)64 GB · 256 GB/s · ~34–52 tok/s$1,959= 67 mo of VPS
  • Mac mini (M5 Pro, 64 GB)64 GB · 307 GB/s · ~40–62 tok/s$2,299= 79 mo of VPS
  • Mac Studio (M5 Max, 36 GB)36 GB · 460 GB/s · ~60–93 tok/s$2,499= 86 mo of VPS
  • Framework Desktop (Ryzen AI Max+ 395, 128 GB)128 GB · 256 GB/s · ~34–52 tok/s$3,449= 119 mo of VPS
  • NVIDIA DGX Spark (128 GB)128 GB · 273 GB/s · ~36–55 tok/s$4,699*= 10+ yrs of VPS

Buy · per month

$37.45

$36.08 hardware + $1.36 power

Rent · per month

$29.05

Contabo Cloud VPS 12 · 12 vCPU · 48 GB

Break-even

The Mac mini (M6, 32 GB) pays for itself after 47 months of replacing the VPS — and it streams an estimated 20–31 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 Qwen3 30B-A3B (MoE) need?
About 20.9 GB for CPU inference at Q4_K_M: 18.6 GB of weights, 819 MB of KV cache at 8,192 tokens of context, and 1.5 GB of headroom. Longer contexts need more KV cache (96 KB per token for this model).
What is the cheapest VPS that can run Qwen3 30B-A3B (MoE)?
Contabo Cloud VPS 12 · 12 vCPU · 48 GB (48 GB RAM, 12 vCPU) at $29.05/mo excl. VAT runs it comfortably as of 5 Sept 2026. Contabo Cloud VPS 8 · 8 vCPU · 24 GB at $16.27/mo is a tight fit.
How fast will Qwen3 30B-A3B (MoE) run on a VPS without a GPU?
Roughly 9–18 tok/s on the top pick. Token generation is bound by memory bandwidth — each token streams the active-expert weights once — so more and dedicated vCPUs help; a GPU or Apple-silicon machine is 10–50× faster.
Same lineupQwen3 0.6BQwen3 1.7BQwen3 4BQwen3 8BQwen3 14BQwen3 32BQwen2.5-Coder 7BQwen2.5-Coder 32BSame size classLlama 4 Scout (109B MoE)gpt-oss-20b (MoE)gpt-oss-120b (MoE)VPS for AI agents (API-hosted models) →