Run Llama 4 Scout (109B MoE) on a VPS.
September 2026 · 2 of 280 plans fit · re-ranked daily
Mixture-of-experts: 109B of weights to hold in RAM but only 17B active per token, so it runs surprisingly fast once it fits. On a CPU-only VPS it needs about 68.4 GB of RAM: 65.4 GB of Q4_K_M weights, 1.5 GB of KV cache for an 8,192-token context and 1.5 GB for the OS and runtime. 2 of the 280 plans in our index fit; the cheapest comfortable pick is Contabo's Cloud VPS Plus 18 · 18 vCPU · 96 GB at $115.06/mo, streaming an estimated 1.2–2.3 tok/s.
RAM needed · CPU inference
68.4 GB
- weights · Q4_K_M
- 65.4 GB
- KV cache · 8K context
- 1.5 GB
- OS + runtime headroom
- 1.5 GB
Other quants: Q8_0 weights 114.5 GB (near-lossless, about half of F16); F16 1.8 GB. Comfortable = 20% headroom over the total.
Model card
- Size
- 109B total · 17B active
- Context
- 10.24M tokens
- Kind
- chat · vision
- Released
- 2025-04
- License
- Llama 4 Community License
Facts fetched from Hugging Face on 3 Sept 2026: exact GGUF file sizes, KV geometry from the GGUF header · 231,450 downloads.
Best VPS plans for Llama 4 Scout (109B MoE)
ranked by estimated tokens/s per dollar, comfortable fits first
- 1runs comfortably~1.2–2.3 tok/s$115.06/mo
- 2runs comfortably~0.7–1.4 tok/s$240/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 · 3 reference machines fit
all machines →- Framework Desktop (Ryzen AI Max+ 395, 128 GB)128 GB · 256 GB/s · ~6.5–10 tok/s$3,449= 30 mo of VPS
- NVIDIA DGX Spark (128 GB)128 GB · 273 GB/s · ~7–11 tok/s$4,699*= 41 mo of VPS
- Mac Studio (M5 Ultra, 96 GB)96 GB · 1200 GB/s · ~31–47 tok/s · tight$5,499= 48 mo of VPS
Buy · per month
$100.47
$95.81 hardware + $4.67 power
Rent · per month
$115.06
Contabo Cloud VPS Plus 18 · 18 vCPU · 96 GB
Break-even
The Framework Desktop (Ryzen AI Max+ 395, 128 GB) pays for itself after 31 months of replacing the VPS — and it streams an estimated 6.5–10 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 Llama 4 Scout (109B MoE) need?
- About 68.4 GB for CPU inference at Q4_K_M: 65.4 GB of weights, 1.5 GB of KV cache at 8,192 tokens of context, and 1.5 GB of headroom. Longer contexts need more KV cache (192 KB per token for this model).
- What is the cheapest VPS that can run Llama 4 Scout (109B MoE)?
- Contabo Cloud VPS Plus 18 · 18 vCPU · 96 GB (96 GB RAM, 18 vCPU) at $115.06/mo excl. VAT runs it comfortably as of 5 Sept 2026.
- How fast will Llama 4 Scout (109B MoE) run on a VPS without a GPU?
- Roughly 1.2–2.3 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.