Run Gemma 3 12B on a VPS.
prices as of · Contabo as of · 178 of 524 plans fit, disk included · re-ranked daily
Mid-size multimodal model; fits a 16 GB plan at Q4 with room for context. On a CPU-only VPS it needs about 11.8 GB of RAM: 7.3 GB of Q4_K_M weights, 3 GB of KV cache for an 8,192-token context and 1.5 GB for the OS and runtime. 178 of the 524 plans in our index fit and include a disk; the cheapest comfortable pick is HOSTKEY's vm.v2-medium · 8 vCPU · 16 GB at $16.15/mo, streaming an estimated 2.6–5.2 tok/s.
The cheapest VPS that runs Gemma 3 12B is HOSTKEY vm.v2-medium at $16.15/mo, 16 GB of RAM against the 11.8 GB the model needs, checked 16 Sept 2026.
Best VPS plans for Gemma 3 12B
ranked by estimated tokens/s per dollar, comfortable fits first
- 1runs comfortably~3.9–7.8 tok/s$20.78/moView at netcup ↗
- 2runs comfortably~2.6–5.2 tok/s$16.15/mochecked 10 days agoView at Contabo ↗
- 3runs comfortably~2.6–5.2 tok/s$16.15/moView at HOSTKEY ↗
- 4runs comfortably~2.6–5.2 tok/s$18.45/monot available nowView at Hetzner ↗
- 5runs comfortably~2.6–5.2 tok/s$18.46/moView at HOSTKEY ↗
- 6runs comfortably~2.6–5.2 tok/s$18.67/moView at netcup ↗
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 neighbors are, so treat these as order-of-magnitude estimates.
RAM needed · CPU inference
11.8 GB
- weights · Q4_K_M
- 7.3 GB
- KV cache · 8K context
- 3 GB
- OS + runtime headroom
- 1.5 GB
Other quants: Q8_0 weights 12.5 GB (near-lossless, about half of F16); F16 23.5 GB. Comfortable = 20% headroom over the total.
Model card
- Size
- 12B parameters
- Context
- 128K tokens
- Kind
- chat · vision
- Released
- 2025-03
- License
- Gemma Terms of Use
Facts fetched from Hugging Face on 7 Sept 2026: exact GGUF file sizes, KV geometry from the GGUF header · 742,412 downloads.
Or rent it as an API · 14 providers sell Gemma 3 12B
every price for Gemma 3 12B →Google list price
$0.30/M tokens
Cheapest price we can explain
none below list
Token prices are per million, three parts input to one part output, and exclude tax. A rented box costs the same whether or not you use it and runs whatever else you put on it; an API charges for what you send and nothing when you stop.
Or buy hardware · 14 reference machines fit
all machines →- Raspberry Pi 5 (16 GB)16 GB · 17 GB/s · ~0.6–0.9 tok/s$305*= 19 mo of VPS
- GeForce RTX 5060 Ti 16 GB (card only)16 GB · 448 GB/s · ~32–49 tok/s$800*= 50 mo of VPS
- Mac mini (M6, 16 GB)16 GB · 153 GB/s · ~11–17 tok/s$899= 56 mo of VPS
- GeForce RTX 5070 Ti 16 GB (card only)16 GB · 896 GB/s · ~64–98 tok/s$1,200*= 74 mo of VPS
- Mac mini (M6, 32 GB)32 GB · 153 GB/s · ~11–17 tok/s$1,299= 80 mo of VPS
- GeForce RTX 5080 16 GB (card only)16 GB · 960 GB/s · ~68–105 tok/s$1,600*= 99 mo of VPS
Buy · per month
$8.78
$8.47 hardware + $0.31 power
Rent · per month
$16.15
HOSTKEY vm.v2-medium · 8 vCPU · 16 GB
Break-even
The Raspberry Pi 5 (16 GB) pays for itself after 19 months of replacing the VPS, and it streams an estimated 0.6–0.9 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 Gemma 3 12B need?
- About 11.8 GB for CPU inference at Q4_K_M: 7.3 GB of weights, 3 GB of KV cache at 8,192 tokens of context, and 1.5 GB of headroom. Longer contexts need more KV cache (384 KB per token for this model).
- What is the cheapest VPS that can run Gemma 3 12B?
- HOSTKEY vm.v2-medium · 8 vCPU · 16 GB (16 GB RAM, 8 vCPU) at $16.15/mo excl. VAT runs it comfortably as of 16 Sept 2026. HOSTKEY vm.medium · 6 vCPU · 12 GB at $11.70/mo is a tight fit.
- How fast will Gemma 3 12B run on a VPS without a GPU?
- Roughly 3.9–7.8 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.