SH1
Sold out
$ 80,730/ mo
- 4 x Intel Xeon
- 4 GB RAM
- 240 GB SSD or 1 TB SAS
- 1 static public IP
- 99.8% network uptime
- Unmetered transfer (fair use)
- Windows or Linux
Open models server
Open models server: your data never leaves your infrastructure
A hybrid cloud server with 8 cores and 32 GB to run open models with Ollama or llama.cpp. It is a CPU server: it suits small quantised models, and we would rather say so before you order.
The short answer
You can run an open model —Hermes, Llama, Qwen— on your own server with Ollama: install it with one command, pull the model and it answers inside the server, with no text going through a third-party API. At SIS Argentina the SH6 (8 cores, 32 GB) comfortably runs quantised 7B to 14B models. With no GPU that means a few tokens per second and little concurrency: enough to classify, summarise or extract data from documents, not enough for a chat for the whole team.
- Ollama
- llama.cpp
- Hermes 3
- Llama 3.1
- Qwen3
- Gemma 3
- Mistral
- DeepSeek-R1
- GGUF
- Ubuntu 24.04 LTS
What you get
Text never leaves the server
Inference runs on your server: no API key, no request to a third party. The only thing that goes out to the internet is the model download, the first time.
No GPU, and it shows
The infrastructure is CPU-based. A quantised 8B model answers at a few tokens per second and handles few concurrent requests. If your case needs a GPU, write to us and we will quote it.
What it is enough for
Classifying text, summarising, extracting data from documents, tagging records and trying a model before deciding. Batch work, where nobody is waiting for the answer on screen.
What it is not for
A chat for the whole team, instant answers or 70B models: that calls for a GPU. Better to know now than after ordering.
32 GB for the model and the rest
The SH6 has 8 cores and 32 GB, with RAID and a spare disk: the model fits in memory and there is room left for the database and the application querying it.
In our Chacarita datacenter
The server sits in our own TIER‑II datacenter in the City of Buenos Aires, with infrastructure in Argentina since 2003.
Open models server plans
These are the usual Cloud Servers plans. Monthly prices, VAT excluded. USD amounts are the official store list; local invoicing is issued in pesos.
SH2
$ 87,458/ mo
- 4 x Intel Xeon
- 6 GB RAM
- 240 GB SSD or 1 TB SAS
- 1 static public IP
- 99.8% network uptime
- Unmetered transfer (fair use)
- Windows or Linux
SH3
$ 94,185/ mo
- 4 x Intel Xeon
- 8 GB RAM
- 240 GB SSD or 1 TB SAS
- 1 static public IP
- 99.8% network uptime
- Unmetered transfer (fair use)
- Windows or Linux
SH4
$ 100,913/ mo
- 8 x Intel Xeon
- 8 GB RAM
- 240 GB SSD or 1 TB SAS
- 1 static public IP
- 99.8% network uptime
- Unmetered transfer (fair use)
- Windows or Linux
SH5
$ 107,640/ mo
- 8 x Intel Xeon
- 16 GB RAM
- 240 GB SSD or 1 TB SAS
- 1 static public IP
- 99.8% network uptime
- Unmetered transfer (fair use)
- Windows or Linux
Most chosen
SH6
$ 114,368/ mo
- 8 x Intel Xeon
- 32 GB RAM
- 240 GB SSD or 1 TB SAS
- 1 static public IP
- 99.8% network uptime
- Unmetered transfer (fair use)
- Windows or Linux
An open model running in three steps
-
Install Ollama
The official script installs it and registers it as a systemd service on Ubuntu 24.04:
curl -fsSL https://ollama.com/install.sh | sh -
Pull a model that fits
Nous Research’s 8B Hermes 3 is a 4.7 GB download and fits easily in 32 GB. Try it right away with ollama run hermes3:8b.
ollama pull hermes3:8b -
Leave it answering inside only
It listens on 127.0.0.1 and nowhere else. Ollama’s local API asks for no username or password: never expose port 11434 to the internet. To query it from another machine, go through an SSH tunnel: ssh -N -L 11434:127.0.0.1:11434 juan@203.0.113.24.
mkdir -p /etc/systemd/system/ollama.service.d && printf '[Service]\nEnvironment="OLLAMA_HOST=127.0.0.1:11434"\n' > /etc/systemd/system/ollama.service.d/red.conf && systemctl daemon-reload && systemctl restart ollama
Read on: Servidor cloud o híbrido: qué hay detrás de cada nombre · Cuánta RAM necesita tu VPS
Frequently asked questions
Which models fit in 32 GB?
Quantised 7B to 14B models fit comfortably: the 8B Hermes 3 is a 4.7 GB download. 70B models do not: hermes3:70b is 40 GB, and even if it fit, without a GPU it would not reach a usable speed.
What speed can I expect without a GPU?
A few tokens per second with a quantised 8B model, and fewer still when it handles several requests at once. Enough for batch work; not enough for someone waiting on the other end. It is worth measuring with your own workload before committing.
Do you have GPU servers?
No: the infrastructure is CPU-based. If your case needs a GPU, write to us and we will quote it.
Can I run Hermes 4?
Hermes 4 weights are published on Hugging Face under the Apache 2.0 licence. What the official Ollama library carries is Hermes 3; for other weights, llama.cpp runs any GGUF file you hand it.
Does any data leave the server?
Inference is local: the text you send the model never leaves the server. The only thing that goes out to the internet is the model download, the first time, plus any updates you choose to pull.
Which plan do I need?
We recommend the SH6: 8 cores and 32 GB, with RAID and a spare disk. If you start with a smaller one to test, you can move up a plan later without migrating or reinstalling.
Keep going
AI agent VPSYour agent running without your laptopCloud ServersHigh-performance machines on Proxmox VE with CPU, memory and network assigned to you alone. Pick Argentina, the USA or Europe with a local IP and scale without migrations or reinstalls.Cloud VPSDedicated resources on Proxmox with RAID 10 SSD storage. Choose your OS, scale vertically or horizontally, and never share what you paid for.