Host open-source LLMs on a dedicated GPU, with your data private.

It is a server with a dedicated NVIDIA GPU, designed to host and run open-source LLMs (Large Language Models) — such as Llama 3, Mistral, DeepSeek, Qwen and Gemma. It comes with Ollama, vLLM and llama.cpp preinstalled. You run inference and, on the Pro plan, fine-tuning, with the GPU 100% yours.

The Inference plan (20 GB GPU) serves 7B to 13B models in solid production — Llama 3 8B, Mistral 7B, Phi-3, Gemma 2. The Pro plan (96 GB GPU) runs large models (Llama 3 70B, Mixtral 8×22B, DeepSeek R1) and enables fine-tuning.

The Inference plan costs US$ 649.80/mo and the Pro US$ 2,575.80/mo. There is a one-time setup fee of US$ 259.80 (it covers preparing the server with the GPU, CUDA drivers and the AI tools). No contract.

Provisioning GPU servers takes up to 48 business hours. Unlike a regular VPS, GPU servers have limited stock and dedicated preparation. The flow is: you request the plan, we confirm availability and delivery, and we provision it.

Yes, completely. The model runs on your server — prompts, responses and training data never leave your infrastructure. That is the fundamental difference from APIs like OpenAI or Anthropic, where all content is sent to third-party servers.

Any open-source LLM: Llama 3, Mistral, Mixtral, DeepSeek, Qwen, Gemma, Phi-3 and others. The Ollama, vLLM and llama.cpp tools come installed. The Pro plan also includes Hugging Face Transformers, Accelerate and PEFT for fine-tuning.

Yes, on the Pro plan (96 GB GPU). It supports LoRA, QLoRA, DPO and DeepSpeed — you adapt an open-source model to your data and domain. The Inference plan is focused on serving models, not training.

It is worth it when volume is high (from ~10 million tokens/month) or data is sensitive (LGPD, healthcare, legal, financial). The cost is fixed (no per-token surprise), data stays in your infrastructure and you swap models without rewriting code. For low volume and non-sensitive data, per-token API is still cheaper.

The LLM Server has a dedicated NVIDIA GPU — high performance for production inference and fine-tuning. The AI Cloud Server runs Ollama on CPU (no GPU), much cheaper, ideal for internal chat, corporate RAG and automations where 8-15 tokens/second is enough.

Ollama and vLLM expose a REST API 100% compatible with OpenAI — just point the SDK to your server URL (e.g. https://your-server.rollin.host/v1) and use it as if it were OpenAI. Open-source models equivalent to GPT-4 (Llama 3 70B, Mixtral 8×22B, DeepSeek R1) run on the Pro plan.

Yes — Rollin Serviços Digitais e Tecnologia LTDA is a Brazilian company with a Tier III datacenter in São Paulo, NF-e billing in BRL and human support 24/7. It is the first Brazilian cloud specialized in AI, with dedicated products for LLM, GPU, vector DB and WhatsApp agents.

Yes — human support 24/7, with people who understand CUDA, Ollama, vLLM and fine-tuning. Rollin Host is a Brazilian company (Rollin Serviços Digitais e Tecnologia LTDA).