Whether you already run a GPU server and want to serve a model on it with vLLM, or you're new and want to dive into private, EU-hosted LLM inference β this one-click Private LLM service is for you.
- One-click deploy: hit Start (hourly or monthly) and we launch a GPU server preinstalled with vLLM using the exact benchmarked config for that model.
- Only what's available: Start is enabled only when the best GPU is actually free to deploy right now.
- Judge quality first: open Show response quality to read real sample answers per test category before you commit.
| Model | Context | Parallel t/s | Single t/s | Hourly | Monthly | Config |
|---|---|---|---|---|---|---|
| Llama 3.3 70B 4-bit Perfect on sparbox.m2 | 12,288 | 68 | 35 | β¬0.79/h | β¬425.00/mo | sparbox.m2 |
| Qwen3 30B 4-bit Perfect on infinityai.s1 | 84,992 | 838 | 153 | β¬0.63/h | β¬340.00/mo | infinityai.s1 |
| Qwen3.6 27B 4-bit Perfect on infinityai.s1 | 58,368 | 402 | 58 | β¬0.63/h | β¬340.00/mo | infinityai.s1 |
| Qwen3.6 35B 4-bit Perfect on sparbox.m2 | 62,464 | 747 | 143 | β¬0.79/h | β¬425.00/mo | sparbox.m2 |
| Gemma 4 31B 4-bit Perfect on sparbox.m2 | 17,408 | 390 | 62 | β¬0.79/h | β¬425.00/mo | sparbox.m2 |
| Gemma 4 E2B Perfect on sparbox.sm1 More context, higher speed on sparbox.m2 | 76,800 | 914 | 133 | β¬0.43/h | β¬230.00/mo | sparbox.sm1 |
| 117,760 +52% | 1,160 +27% | 184 | β¬0.79/h | β¬425.00/mo | sparbox.m2 | |
| Gemma 4 E4B Perfect on sparbox.sm1 More context, higher speed on sparbox.m2 | 30,720 | 538 | 70 | β¬0.43/h | β¬230.00/mo | sparbox.sm1 |
| 95,232 +205% | 737 +37% | 104 | β¬0.79/h | β¬425.00/mo | sparbox.m2 | |
| Granite 4.1 8B Perfect on infinityai.s1 | 69,632 | 539 | 69 | β¬0.63/h | β¬340.00/mo | infinityai.s1 |
| Llama 3.1 8B Perfect on infinityai.s1 | 76,800 | 597 | 78 | β¬0.63/h | β¬340.00/mo | infinityai.s1 |
| Phi 4 Perfect on infinityai.s1 | 14,336 | 339 | 44 | β¬0.63/h | β¬340.00/mo | infinityai.s1 |
| Phi 4 multimodal Perfect on infinityai.s1 | 105,472 | 977 | 132 | β¬0.63/h | β¬340.00/mo | infinityai.s1 |
| Ministral 3 14B Perfect on sparbox.sm1 More context, higher speed on sparbox.m2 | 20,480 | 429 | 56 | β¬0.43/h | β¬230.00/mo | sparbox.sm1 |
| 84,992 +300% | 623 +45% | 89 | β¬0.79/h | β¬425.00/mo | sparbox.m2 | |
| Ministral 3 3B Perfect on sparbox.sm1 | 84,992 | 1,218 | 164 | β¬0.43/h | β¬230.00/mo | sparbox.sm1 |
| Mistral 7B Perfect on infinityai.s1 | 28,672 | 631 | 82 | β¬0.63/h | β¬340.00/mo | infinityai.s1 |
| Llama 3.1 8B 8-bit Perfect on infinityai.s1 | 117,760 | 889 | 115 | β¬0.63/h | β¬340.00/mo | infinityai.s1 |
| GPT OSS 20B Perfect on infinityai.s1 | 117,760 | 685 | 201 | β¬0.63/h | β¬340.00/mo | infinityai.s1 |
| Ornith 1.0 9B 4-bit Perfect on ranger.s1 | 52,224 | 78 | 59 | β¬0.41/h | β¬220.00/mo | ranger.s1 |
| Qwen3 Coder 30B 4-bit Perfect on sparbox.sm1 More context, higher speed on sparbox.m2 | 25,600 | 886 | 170 | β¬0.43/h | β¬230.00/mo | sparbox.sm1 |
| 84,992 +224% | 949 +7% | 167 | β¬0.79/h | β¬425.00/mo | sparbox.m2 | |
| Qwen3 VL 32B 4-bit Perfect on infinityai.s1 | 28,672 | 421 | 56 | β¬0.63/h | β¬340.00/mo | infinityai.s1 |
| Qwen2.5 32B 4-bit Perfect on infinityai.s1 | 28,672 | 429 | 57 | β¬0.63/h | β¬340.00/mo | infinityai.s1 |
| Qwen2.5 7B Perfect on infinityai.s1 | 28,672 | 628 | 80 | β¬0.63/h | β¬340.00/mo | infinityai.s1 |
| Qwen3 14B Perfect on infinityai.s1 | 25,600 | 342 | 44 | β¬0.63/h | β¬340.00/mo | infinityai.s1 |
| Qwen3 32B 8-bit Perfect on sparbox.m2 | 18,432 | 297 | 42 | β¬0.79/h | β¬425.00/mo | sparbox.m2 |
| Qwen3 4B Perfect on sparbox.sm1 | 36,864 | 627 | 85 | β¬0.43/h | β¬230.00/mo | sparbox.sm1 |
| Qwen3 4B (2507) Perfect on infinityai.s1 | 117,760 | 918 | 119 | β¬0.63/h | β¬340.00/mo | infinityai.s1 |
| Qwen3 8B Perfect on infinityai.s1 | 36,864 | 590 | 75 | β¬0.63/h | β¬340.00/mo | infinityai.s1 |
| Qwen3 Coder 30B 8-bit Perfect on infinityai.s1 More context on sparbox.m2 | 25,600 | 777 | 136 | β¬0.63/h | β¬340.00/mo | infinityai.s1 |
| 41,984 +62% | 812 | 153 | β¬0.79/h | β¬425.00/mo | sparbox.m2 | |
| Qwen3 VL 32B 8-bit Perfect on sparbox.m2 | 15,360 | 294 | 41 | β¬0.79/h | β¬425.00/mo | sparbox.m2 |
| Qwen3 VL 4B Perfect on infinityai.s1 | 84,992 | 893 | 117 | β¬0.63/h | β¬340.00/mo | infinityai.s1 |
| Qwen3.5 9B Perfect on sparbox.m2 | 95,232 | 558 | 82 | β¬0.79/h | β¬425.00/mo | sparbox.m2 |
| Qwen3.6 27B 8-bit Perfect on infinityai.s1 | 23,552 | 291 | 41 | β¬0.63/h | β¬340.00/mo | infinityai.s1 |
| Gemma 4 31B 8-bit Perfect on sparbox.m2 | 8,192 | 286 | 43 | β¬0.79/h | β¬425.00/mo | sparbox.m2 |
Context = max usable context window. Parallel/Single = tokens/sec. Pricing reflects the best price-performance GPU per model. Prices are prepaid; you keep full root access.
Just want the GPU server?
You don't have to pick a model here β order a bare GPU server and run whatever you like on it, with full root access.
Order a GPU ServerYour vLLM dashboard, ready right after deployment
Seconds after you hit Start, your private model is live behind an OpenAI-compatible endpoint β monitor throughput, context usage and requests from your own dashboard.
Top 10 features of your private LLM GPU server
- One-click vLLM deploy β your model boots with the exact benchmarked config (context length, parallelism, launch flags)
- OpenAI-compatible API out of the box β point your existing SDK at your private endpoint
- Fully private inference β your prompts and data never leave your own server
- Tier-III (NL/FR) & ISO 27001 (DE) EU data centers πͺπΊ, GDPR-compliant with one-click DPA/AVV
- Full root SSH access β fine-tune, swap models or install extra AI software alongside vLLM
- Professionally tested GPU drivers & latest CUDA β no driver/version wrangling
- 100% bare-metal GPU performance β the tokens/sec you see benchmarked are the tokens/sec you get
- Free SSL π on your endpoint (HTTPS & WSS) plus a customizable network-level firewall
- Persistent storage with pause/freeze β keep your model & weights, stop paying while idle
- π Upgrade hardware anytime β move to a bigger GPU for larger context or models without reinstalling
Managed GPU server preinstalled with vLLM
Trooper.AI gives you a managed GPU server preinstalled with vLLM, ready to serve any open large language model through a fast, OpenAI-compatible API. Instead of spending hours picking a GPU, installing CUDA drivers, compiling vLLM and tuning launch flags, you pick a model above and click Start. Within minutes you get a dedicated EU-hosted GPU server with vLLM already running the exact configuration we benchmarked for that model β context length, parallelism and command-line arguments included.
Why a preinstalled vLLM server?
vLLM is the industry-standard high-throughput inference engine, but getting it production-ready is fiddly: matching driver and CUDA versions, choosing --max-num-batched-tokens, --max-num-seqs and the right context window for your GPU's VRAM, and validating that the model actually loads. Our managed GPU server preinstalled with vLLM removes that work. Every configuration on this page comes straight from automated benchmarks on the real hardware, so the server you deploy behaves exactly like the tested run β no trial and error.
Private, EU-hosted and yours
Each server runs on bare-metal GPUs in ISO/IEC 27001-certified, GDPR-compliant German data centers. You get full root SSH access, a persistent machine, and a private endpoint β your prompts and data never leave your server. Because it is a full GPU server (not shared inference), you can also install additional AI software, fine-tune, or run image and audio models alongside your LLM.
Transparent price-performance
For every model we show the GPU with the best price per performance, with honest hourly and monthly pricing. You can review real sample answers per model via Show response quality before deploying, so you know both the speed and the answer quality you are paying for. Pay hourly to experiment or monthly for production β a managed GPU server preinstalled with vLLM that scales with your needs.
Your Own Private LLM GPU Server
A dedicated private LLM GPU server gives you the whole machine: full root access, a private OpenAI-compatible endpoint, unlimited requests at a flat prepaid rate, and the freedom to fine-tune, swap models or run image and audio workloads alongside your LLM. Because the GPU is yours, throughput and context length are predictable and your prompts never leave your server β ideal for steady traffic and strict privacy requirements.
Pick a model from the matrix above and click Start: within minutes you get an EU-hosted GPU server with vLLM already running the exact benchmarked configuration β context length, parallelism and launch flags included. Pay hourly to experiment or monthly for production, and upgrade to a bigger GPU anytime for larger context or models without reinstalling.
Reliable Hardware, Built by Experts
Behind every private LLM GPU server is enterprise-grade, upcycled hardware maintained by our own team. Here, Markus and Jaimie are racking an NVIDIA A100 cluster in one of our ISO/IEC 27001-certified colocation data centers in Germany β the same class of GPU servers you deploy from this page. We upcycle high-performance components into optimized inference rigs, extending hardware lifecycles while reducing e-waste. We don't resell third-party capacity; we own and operate our own hardware in colocation data centers in Germany and the Netherlands, so we can guarantee performance, security, and data residency at every layer of the stack.
OpenAI Chat Completions API Compatible β Migrate Your AI Stack in Minutes
Your private vLLM server exposes an endpoint that is 100% compatible with the OpenAI Chat Completions API format (/v1/chat/completions). If your application already uses the OpenAI SDK β Python, Node.js, or any HTTP client β pointing it at your own server is a one-line change: update the base URL and API key. You get the same request and response schema, and full support for streaming, JSON mode, function calling, and multimodal inputs. No code rewrite, no new abstractions, no vendor lock-in β your integration stays portable and you stay in control.
Looking for an OpenAI API alternative hosted in Europe? A private LLM GPU server gives you equivalent Chat Completions API functionality with EU data residency, a flat prepaid price, and full ownership of the machine.
Works with the Tools You Already Use
Because your server speaks the OpenAI Chat Completions API, you can plug it into virtually any AI tool, IDE or automation platform β just point the base URL at your own endpoint and add your key. A few popular examples:
β¦ and many more β any app, agent or SDK that supports an OpenAI-compatible endpoint works out of the box.